ModelVerdict

Leaderboard

Which model does your task best

Sorted by the share of correct fields. Cost is the real token spend scaled to 1,000 documents. Bar: 95% correct fields; below it, manual fixes cost more than a cheaper model saves.

Every model gets the same prompt, with no per-model tuning. What that means →

Compare two models side by side →

Invoice extraction · Czech · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5 (Anthropic) — 100.00% · $9.41 / 1000Gemini 3.8 Flash (Google) — 99.95% · $4.26 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $10.79 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.90% · $0.50 / 1000Claude Haiku 5.5 (Anthropic) — 99.90% · $0.62 / 1000GPT-6 Sol (OpenAI) — 99.90% · $10.12 / 1000GLM-5.3 Flash (Z.ai) — 99.69% · $1.02 / 1000Gemini 3.5 Flash Lite (Google) — 99.49% · $1.03 / 1000GPT-6 Luna (OpenAI) — 99.28% · $0.51 / 1000Mistral Large 4 (Mistral) — 97.08% · $5.69 / 1000Gemma 4 31B (Google) — 96.10% · $0.21 / 1000Mistral Small 4 (Mistral) — 93.64% · $0.44 / 1000Mistral Medium 3.5 (Mistral) — 93.49% · $4.69 / 1000Qwen3.8 Flash (Qwen) — 91.79% · $1.02 / 1000Gemma 4 31BDeepSeek V4.1 FlashGemini 3.8 FlashClaude Sonnet 5100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5 (Anthropic) — 100.00% · $9.41 / 1000Gemini 3.8 Flash (Google) — 99.95% · $4.26 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $10.79 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.90% · $0.50 / 1000Claude Haiku 5.5 (Anthropic) — 99.90% · $0.62 / 1000GPT-6 Sol (OpenAI) — 99.90% · $10.12 / 1000GLM-5.3 Flash (Z.ai) — 99.69% · $1.02 / 1000Gemini 3.5 Flash Lite (Google) — 99.49% · $1.03 / 1000GPT-6 Luna (OpenAI) — 99.28% · $0.51 / 1000Mistral Large 4 (Mistral) — 97.08% · $5.69 / 1000Gemma 4 31B (Google) — 96.10% · $0.21 / 1000Mistral Small 4 (Mistral) — 93.64% · $0.44 / 1000Mistral Medium 3.5 (Mistral) — 93.49% · $4.69 / 1000Qwen3.8 Flash (Qwen) — 91.79% · $1.02 / 1000Gemma 4 31BDeepSeek V4.1 FlashGemini 3.8 FlashClaude Sonnet 5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5
AnthropicEU data
100.00%$9.4103.6 sec
2Gemini 3.8 Flash
Google
99.95%$4.2673.5 sec
3Claude Sonnet 5.5New
AnthropicEU data
99.95%$10.7974.4 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
99.90%$0.50135.0 sec
5Claude Haiku 5.5New
AnthropicEU data
99.90%$0.62133.4 sec
6GPT-6 Sol
OpenAIEU data
99.90%$10.12133.9 sec
7GLM-5.3 FlashNew
Z.ai · open weights
99.69%$1.02334.5 sec
8Gemini 3.5 Flash Lite
GoogleEU data
99.49%$1.03671.8 sec
9GPT-6 Luna
OpenAIEU data
99.28%$0.51673.2 sec
10Mistral Large 4New
MistralEU data
97.08%$5.6911326.3 sec
11Gemma 4 31B
Google · open weights
96.10%$0.2133310.4 sec
12Mistral Small 4
Mistral · open weightsEU data
93.64%$0.444402.0 sec
13Mistral Medium 3.5
MistralEU data
93.49%$4.693672.6 sec
14Qwen3.8 Flash
Qwen
91.79%$1.0223317.0 sec

Invoice extraction · Czech · PDF text (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.21 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.37 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.50 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.59 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.91 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.46 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.02 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.42 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.87 / 1000Mistral Medium 3.5 (Mistral) — 99.83% · $2.97 / 1000Claude Sonnet 5.5 (Anthropic) — 99.83% · $6.32 / 1000Gemma 4 31B (Google) — 99.49% · $0.24 / 1000Mistral Small 4 (Mistral) — 96.75% · $0.26 / 1000Qwen3.8 Flash (Qwen) — 91.97% · $0.74 / 1000GPT-6 Luna100%95%90%85%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.21 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.37 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.50 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.59 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.91 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.46 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.02 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.42 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.87 / 1000Mistral Medium 3.5 (Mistral) — 99.83% · $2.97 / 1000Claude Sonnet 5.5 (Anthropic) — 99.83% · $6.32 / 1000Gemma 4 31B (Google) — 99.49% · $0.24 / 1000Mistral Small 4 (Mistral) — 96.75% · $0.26 / 1000Qwen3.8 Flash (Qwen) — 91.97% · $0.74 / 1000GPT-6 Luna
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GPT-6 Luna
OpenAIEU data
100.00%$0.2102.6 sec
2Claude Haiku 5.5New
AnthropicEU data
100.00%$0.3702.3 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.5006.0 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.5904.8 sec
5Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.9101.1 sec
6Gemini 3.8 Flash
Google
100.00%$3.4602.8 sec
7GPT-6 Sol
OpenAIEU data
100.00%$4.0202.8 sec
8Mistral Large 4New
MistralEU data
100.00%$4.42021.2 sec
9Claude Sonnet 5
AnthropicEU data
100.00%$4.8702.7 sec
10Mistral Medium 3.5
MistralEU data
99.83%$2.97221.5 sec
11Claude Sonnet 5.5New
AnthropicEU data
99.83%$6.32224.5 sec
12Gemma 4 31B
Google · open weights
99.49%$0.24679.0 sec
13Mistral Small 4
Mistral · open weightsEU data
96.75%$0.262441.4 sec
14Qwen3.8 Flash
Qwen
91.97%$0.7446719.9 sec

Invoice extraction · Czech · Clean image (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.49 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.61 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $13.20 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $14.28 / 1000GPT-6 Sol (OpenAI) — 100.00% · $14.57 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.82 / 1000GLM-5.3 Flash (Z.ai) — 99.83% · $1.39 / 1000GPT-6 Luna (OpenAI) — 99.66% · $0.73 / 1000Gemini 3.5 Flash Lite (Google) — 99.49% · $1.07 / 1000Mistral Large 4 (Mistral) — 96.24% · $5.85 / 1000Gemma 4 31B (Google) — 95.38% · $0.20 / 1000Qwen3.8 Flash (Qwen) — 92.65% · $1.11 / 1000Mistral Small 4 (Mistral) — 92.31% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 92.14% · $5.37 / 1000Gemma 4 31BDeepSeek V4.1 Flash100%95%90%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.49 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.61 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $13.20 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $14.28 / 1000GPT-6 Sol (OpenAI) — 100.00% · $14.57 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.82 / 1000GLM-5.3 Flash (Z.ai) — 99.83% · $1.39 / 1000GPT-6 Luna (OpenAI) — 99.66% · $0.73 / 1000Gemini 3.5 Flash Lite (Google) — 99.49% · $1.07 / 1000Mistral Large 4 (Mistral) — 96.24% · $5.85 / 1000Gemma 4 31B (Google) — 95.38% · $0.20 / 1000Qwen3.8 Flash (Qwen) — 92.65% · $1.11 / 1000Mistral Small 4 (Mistral) — 92.31% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 92.14% · $5.37 / 1000Gemma 4 31BDeepSeek V4.1 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.4904.4 sec
2Gemini 3.8 Flash
Google
100.00%$4.6104.0 sec
3Claude Sonnet 5
AnthropicEU data
100.00%$13.2003.8 sec
4Claude Sonnet 5.5New
AnthropicEU data
100.00%$14.2804.3 sec
5GPT-6 Sol
OpenAIEU data
100.00%$14.5703.9 sec
6Claude Haiku 5.5New
AnthropicEU data
99.83%$0.82223.9 sec
7GLM-5.3 FlashNew
Z.ai · open weights
99.83%$1.39224.6 sec
8GPT-6 Luna
OpenAIEU data
99.66%$0.73443.6 sec
9Gemini 3.5 Flash Lite
GoogleEU data
99.49%$1.07672.0 sec
10Mistral Large 4New
MistralEU data
96.24%$5.8515622.9 sec
11Gemma 4 31B
Google · open weights
95.38%$0.2037811.0 sec
12Qwen3.8 Flash
Qwen
92.65%$1.1115615.9 sec
13Mistral Small 4
Mistral · open weightsEU data
92.31%$0.515561.9 sec
14Mistral Medium 3.5
MistralEU data
92.14%$5.374002.5 sec

Invoice extraction · Czech · Scan (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5 (Anthropic) — 100.00% · $10.49 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $11.99 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.68 / 1000Gemini 3.8 Flash (Google) — 99.83% · $4.61 / 1000GPT-6 Sol (OpenAI) — 99.83% · $11.89 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.66% · $0.49 / 1000GLM-5.3 Flash (Z.ai) — 99.32% · $1.14 / 1000Gemini 3.5 Flash Lite (Google) — 98.97% · $1.08 / 1000GPT-6 Luna (OpenAI) — 98.29% · $0.59 / 1000Gemma 4 31B (Google) — 94.87% · $0.20 / 1000Mistral Large 4 (Mistral) — 94.53% · $6.67 / 1000Mistral Small 4 (Mistral) — 92.14% · $0.51 / 1000Qwen3.8 Flash (Qwen) — 90.94% · $1.17 / 1000Mistral Medium 3.5 (Mistral) — 89.06% · $5.41 / 1000Gemma 4 31BDeepSeek V4.1 FlashClaude Haiku 5.5Claude Sonnet 5100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5 (Anthropic) — 100.00% · $10.49 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $11.99 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.68 / 1000Gemini 3.8 Flash (Google) — 99.83% · $4.61 / 1000GPT-6 Sol (OpenAI) — 99.83% · $11.89 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.66% · $0.49 / 1000GLM-5.3 Flash (Z.ai) — 99.32% · $1.14 / 1000Gemini 3.5 Flash Lite (Google) — 98.97% · $1.08 / 1000GPT-6 Luna (OpenAI) — 98.29% · $0.59 / 1000Gemma 4 31B (Google) — 94.87% · $0.20 / 1000Mistral Large 4 (Mistral) — 94.53% · $6.67 / 1000Mistral Small 4 (Mistral) — 92.14% · $0.51 / 1000Qwen3.8 Flash (Qwen) — 90.94% · $1.17 / 1000Mistral Medium 3.5 (Mistral) — 89.06% · $5.41 / 1000Gemma 4 31BClaude Haiku 5.5Claude Sonnet 5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5
AnthropicEU data
100.00%$10.4903.7 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$11.9904.4 sec
3Claude Haiku 5.5New
AnthropicEU data
99.83%$0.68223.6 sec
4Gemini 3.8 Flash
Google
99.83%$4.61223.5 sec
5GPT-6 Sol
OpenAIEU data
99.83%$11.89224.2 sec
6DeepSeek V4.1 Flash
DeepSeek · open weights
99.66%$0.49444.8 sec
7GLM-5.3 FlashNew
Z.ai · open weights
99.32%$1.14674.3 sec
8Gemini 3.5 Flash Lite
GoogleEU data
98.97%$1.081331.8 sec
9GPT-6 Luna
OpenAIEU data
98.29%$0.591333.4 sec
10Gemma 4 31B
Google · open weights
94.87%$0.2044411.9 sec
11Mistral Large 4New
MistralEU data
94.53%$6.6715630.5 sec
12Mistral Small 4
Mistral · open weightsEU data
92.14%$0.515332.6 sec
13Qwen3.8 Flash
Qwen
90.94%$1.1711115.4 sec
14Mistral Medium 3.5
MistralEU data
89.06%$5.416442.9 sec

Invoice extraction · Czech · Photo (15 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.54 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.60 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.53 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $8.45 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $10.14 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $0.90 / 1000Gemini 3.5 Flash Lite (Google) — 99.49% · $1.11 / 1000GPT-6 Sol (OpenAI) — 99.49% · $9.76 / 1000GPT-6 Luna (OpenAI) — 98.97% · $0.47 / 1000Mistral Large 4 (Mistral) — 98.46% · $6.03 / 1000Mistral Small 4 (Mistral) — 92.82% · $0.54 / 1000Mistral Medium 3.5 (Mistral) — 91.79% · $5.65 / 1000Gemma 4 31B (Google) — 91.79% · $0.20 / 1000Qwen3.8 Flash (Qwen) — 91.28% · $1.11 / 1000Gemma 4 31BGPT-6 LunaDeepSeek V4.1 Flash100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.54 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.60 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.53 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $8.45 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $10.14 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $0.90 / 1000Gemini 3.5 Flash Lite (Google) — 99.49% · $1.11 / 1000GPT-6 Sol (OpenAI) — 99.49% · $9.76 / 1000GPT-6 Luna (OpenAI) — 98.97% · $0.47 / 1000Mistral Large 4 (Mistral) — 98.46% · $6.03 / 1000Mistral Small 4 (Mistral) — 92.82% · $0.54 / 1000Mistral Medium 3.5 (Mistral) — 91.79% · $5.65 / 1000Gemma 4 31B (Google) — 91.79% · $0.20 / 1000Qwen3.8 Flash (Qwen) — 91.28% · $1.11 / 1000Gemma 4 31BGPT-6 LunaDeepSeek V4.1 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.5405.4 sec
2Claude Haiku 5.5New
AnthropicEU data
100.00%$0.6003.4 sec
3Gemini 3.8 Flash
Google
100.00%$4.5303.4 sec
4Claude Sonnet 5
AnthropicEU data
100.00%$8.4503.9 sec
5Claude Sonnet 5.5New
AnthropicEU data
100.00%$10.1404.3 sec
6GLM-5.3 FlashNew
Z.ai · open weights
99.49%$0.90674.5 sec
7Gemini 3.5 Flash Lite
GoogleEU data
99.49%$1.11671.8 sec
8GPT-6 Sol
OpenAIEU data
99.49%$9.76674.9 sec
9GPT-6 Luna
OpenAIEU data
98.97%$0.471333.3 sec
10Mistral Large 4New
MistralEU data
98.46%$6.0320030.9 sec
11Mistral Small 4
Mistral · open weightsEU data
92.82%$0.544002.9 sec
12Mistral Medium 3.5
MistralEU data
91.79%$5.654673.0 sec
13Gemma 4 31B
Google · open weights
91.79%$0.2066711.0 sec
14Qwen3.8 Flash
Qwen
91.28%$1.1113315.4 sec

Invoice extraction · German · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.5 Flash Lite (Google) — 100.00% · $1.03 / 1000GPT-6 Sol (OpenAI) — 99.95% · $10.65 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $11.21 / 1000Claude Haiku 5.5 (Anthropic) — 99.85% · $0.71 / 1000GLM-5.3 Flash (Z.ai) — 99.54% · $1.01 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.54% · $0.57 / 1000GPT-6 Luna (OpenAI) — 99.23% · $0.54 / 1000Claude Sonnet 5 (Anthropic) — 98.97% · $10.42 / 1000Mistral Large 4 (Mistral) — 97.49% · $6.72 / 1000Mistral Medium 3.5 (Mistral) — 97.38% · $4.72 / 1000Gemini 3.8 Flash (Google) — 97.23% · $7.15 / 1000Gemma 4 31B (Google) — 94.87% · $0.20 / 1000Mistral Small 4 (Mistral) — 94.82% · $0.44 / 1000Qwen3.8 Flash (Qwen) — 93.49% · $1.12 / 1000Gemma 4 31BGPT-6 LunaDeepSeek V4.1 FlashGemini 3.5 Flash Lite100%95%90%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.5 Flash Lite (Google) — 100.00% · $1.03 / 1000GPT-6 Sol (OpenAI) — 99.95% · $10.65 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $11.21 / 1000Claude Haiku 5.5 (Anthropic) — 99.85% · $0.71 / 1000GLM-5.3 Flash (Z.ai) — 99.54% · $1.01 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.54% · $0.57 / 1000GPT-6 Luna (OpenAI) — 99.23% · $0.54 / 1000Claude Sonnet 5 (Anthropic) — 98.97% · $10.42 / 1000Mistral Large 4 (Mistral) — 97.49% · $6.72 / 1000Mistral Medium 3.5 (Mistral) — 97.38% · $4.72 / 1000Gemini 3.8 Flash (Google) — 97.23% · $7.15 / 1000Gemma 4 31B (Google) — 94.87% · $0.20 / 1000Mistral Small 4 (Mistral) — 94.82% · $0.44 / 1000Qwen3.8 Flash (Qwen) — 93.49% · $1.12 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5Gemini 3.5 Flash Lite
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.5 Flash Lite
GoogleEU data
100.00%$1.0301.5 sec
2GPT-6 Sol
OpenAIEU data
99.95%$10.6574.9 sec
3Claude Sonnet 5.5New
AnthropicEU data
99.95%$11.2176.0 sec
4Claude Haiku 5.5New
AnthropicEU data
99.85%$0.71203.9 sec
5GLM-5.3 FlashNew
Z.ai · open weights
99.54%$1.01534.4 sec
6DeepSeek V4.1 Flash
DeepSeek · open weights
99.54%$0.57606.4 sec
7GPT-6 Luna
OpenAIEU data
99.23%$0.54873.7 sec
8Claude Sonnet 5
AnthropicEU data
98.97%$10.421333.8 sec
9Mistral Large 4New
MistralEU data
97.49%$6.728735.3 sec
10Mistral Medium 3.5
MistralEU data
97.38%$4.722532.5 sec
11Gemini 3.8 Flash
Google
97.23%$7.15404.7 sec
12Gemma 4 31B
Google · open weights
94.87%$0.203607.7 sec
13Mistral Small 4
Mistral · open weightsEU data
94.82%$0.444872.0 sec
14Qwen3.8 Flash
Qwen
93.49%$1.1212719.0 sec

Invoice extraction · German · PDF text (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.22 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.45 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.49 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.91 / 1000Qwen3.8 Flash (Qwen) — 100.00% · $0.99 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.85 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.38 / 1000GLM-5.3 Flash (Z.ai) — 99.83% · $0.52 / 1000Gemma 4 31B (Google) — 99.49% · $0.23 / 1000Mistral Medium 3.5 (Mistral) — 99.49% · $2.96 / 1000Claude Sonnet 5 (Anthropic) — 99.32% · $5.89 / 1000Mistral Small 4 (Mistral) — 97.95% · $0.25 / 1000Gemini 3.8 Flash (Google) — 97.78% · $6.24 / 1000Mistral Large 4 (Mistral) — 97.44% · $5.89 / 1000GPT-6 Luna100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.22 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.45 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.49 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.91 / 1000Qwen3.8 Flash (Qwen) — 100.00% · $0.99 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.85 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.38 / 1000GLM-5.3 Flash (Z.ai) — 99.83% · $0.52 / 1000Gemma 4 31B (Google) — 99.49% · $0.23 / 1000Mistral Medium 3.5 (Mistral) — 99.49% · $2.96 / 1000Claude Sonnet 5 (Anthropic) — 99.32% · $5.89 / 1000Mistral Small 4 (Mistral) — 97.95% · $0.25 / 1000Gemini 3.8 Flash (Google) — 97.78% · $6.24 / 1000Mistral Large 4 (Mistral) — 97.44% · $5.89 / 1000GPT-6 Luna
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GPT-6 Luna
OpenAIEU data
100.00%$0.2203.0 sec
2Claude Haiku 5.5New
AnthropicEU data
100.00%$0.4502.8 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.4906.7 sec
4Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.9101.1 sec
5Qwen3.8 Flash
Qwen
100.00%$0.99027.8 sec
6GPT-6 Sol
OpenAIEU data
100.00%$3.8503.1 sec
7Claude Sonnet 5.5New
AnthropicEU data
100.00%$6.3803.9 sec
8GLM-5.3 FlashNew
Z.ai · open weights
99.83%$0.52224.3 sec
9Gemma 4 31B
Google · open weights
99.49%$0.23677.0 sec
10Mistral Medium 3.5
MistralEU data
99.49%$2.96671.4 sec
11Claude Sonnet 5
AnthropicEU data
99.32%$5.89893.3 sec
12Mistral Small 4
Mistral · open weightsEU data
97.95%$0.252001.3 sec
13Gemini 3.8 Flash
Google
97.78%$6.24223.9 sec
14Mistral Large 4New
MistralEU data
97.44%$5.896733.1 sec

Invoice extraction · German · Clean image (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10$20Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.5 Flash Lite (Google) — 100.00% · $1.07 / 1000GPT-6 Sol (OpenAI) — 100.00% · $15.26 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.87 / 1000Claude Sonnet 5.5 (Anthropic) — 99.83% · $13.98 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.66% · $0.55 / 1000GPT-6 Luna (OpenAI) — 99.66% · $0.77 / 1000GLM-5.3 Flash (Z.ai) — 99.15% · $1.33 / 1000Claude Sonnet 5 (Anthropic) — 98.97% · $13.27 / 1000Mistral Large 4 (Mistral) — 97.09% · $6.87 / 1000Mistral Medium 3.5 (Mistral) — 96.58% · $5.43 / 1000Gemini 3.8 Flash (Google) — 95.21% · $8.81 / 1000Gemma 4 31B (Google) — 93.85% · $0.19 / 1000Mistral Small 4 (Mistral) — 93.68% · $0.51 / 1000Qwen3.8 Flash (Qwen) — 88.38% · $1.26 / 1000Gemma 4 31BDeepSeek V4.1 FlashClaude Haiku 5.5Gemini 3.5 Flash Lite100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.5 Flash Lite (Google) — 100.00% · $1.07 / 1000GPT-6 Sol (OpenAI) — 100.00% · $15.26 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.87 / 1000Claude Sonnet 5.5 (Anthropic) — 99.83% · $13.98 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.66% · $0.55 / 1000GPT-6 Luna (OpenAI) — 99.66% · $0.77 / 1000GLM-5.3 Flash (Z.ai) — 99.15% · $1.33 / 1000Claude Sonnet 5 (Anthropic) — 98.97% · $13.27 / 1000Mistral Large 4 (Mistral) — 97.09% · $6.87 / 1000Mistral Medium 3.5 (Mistral) — 96.58% · $5.43 / 1000Gemini 3.8 Flash (Google) — 95.21% · $8.81 / 1000Gemma 4 31B (Google) — 93.85% · $0.19 / 1000Mistral Small 4 (Mistral) — 93.68% · $0.51 / 1000Qwen3.8 Flash (Qwen) — 88.38% · $1.26 / 1000Gemma 4 31BClaude Haiku 5.5Gemini 3.5 Flash Lite
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.5 Flash Lite
GoogleEU data
100.00%$1.0701.7 sec
2GPT-6 Sol
OpenAIEU data
100.00%$15.2605.6 sec
3Claude Haiku 5.5New
AnthropicEU data
99.83%$0.87224.3 sec
4Claude Sonnet 5.5New
AnthropicEU data
99.83%$13.98226.0 sec
5DeepSeek V4.1 Flash
DeepSeek · open weights
99.66%$0.55445.7 sec
6GPT-6 Luna
OpenAIEU data
99.66%$0.77444.1 sec
7GLM-5.3 FlashNew
Z.ai · open weights
99.15%$1.33894.5 sec
8Claude Sonnet 5
AnthropicEU data
98.97%$13.271333.8 sec
9Mistral Large 4New
MistralEU data
97.09%$6.8711134.3 sec
10Mistral Medium 3.5
MistralEU data
96.58%$5.433112.5 sec
11Gemini 3.8 Flash
Google
95.21%$8.81895.1 sec
12Gemma 4 31B
Google · open weights
93.85%$0.194897.4 sec
13Mistral Small 4
Mistral · open weightsEU data
93.68%$0.516221.8 sec
14Qwen3.8 Flash
Qwen
88.38%$1.2617816.7 sec

Invoice extraction · German · Scan (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.5 Flash Lite (Google) — 100.00% · $1.08 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $13.49 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.81 / 1000GPT-6 Sol (OpenAI) — 99.83% · $13.15 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $1.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.97% · $0.61 / 1000Claude Sonnet 5 (Anthropic) — 98.63% · $11.88 / 1000GPT-6 Luna (OpenAI) — 98.29% · $0.66 / 1000Gemini 3.8 Flash (Google) — 97.78% · $6.65 / 1000Mistral Large 4 (Mistral) — 97.09% · $7.10 / 1000Qwen3.8 Flash (Qwen) — 96.92% · $1.15 / 1000Mistral Medium 3.5 (Mistral) — 96.92% · $5.45 / 1000Mistral Small 4 (Mistral) — 92.99% · $0.52 / 1000Gemma 4 31B (Google) — 91.79% · $0.19 / 1000Gemma 4 31BMistral Small 4DeepSeek V4.1 FlashClaude Haiku 5.5Gemini 3.5 Flash Lite100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.5 Flash Lite (Google) — 100.00% · $1.08 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $13.49 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.81 / 1000GPT-6 Sol (OpenAI) — 99.83% · $13.15 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $1.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.97% · $0.61 / 1000Claude Sonnet 5 (Anthropic) — 98.63% · $11.88 / 1000GPT-6 Luna (OpenAI) — 98.29% · $0.66 / 1000Gemini 3.8 Flash (Google) — 97.78% · $6.65 / 1000Mistral Large 4 (Mistral) — 97.09% · $7.10 / 1000Qwen3.8 Flash (Qwen) — 96.92% · $1.15 / 1000Mistral Medium 3.5 (Mistral) — 96.92% · $5.45 / 1000Mistral Small 4 (Mistral) — 92.99% · $0.52 / 1000Gemma 4 31B (Google) — 91.79% · $0.19 / 1000Gemma 4 31BMistral Small 4Claude Haiku 5.5Gemini 3.5 Flash Lite
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.5 Flash Lite
GoogleEU data
100.00%$1.0801.6 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$13.4906.4 sec
3Claude Haiku 5.5New
AnthropicEU data
99.83%$0.81224.2 sec
4GPT-6 Sol
OpenAIEU data
99.83%$13.15225.1 sec
5GLM-5.3 FlashNew
Z.ai · open weights
99.49%$1.19674.4 sec
6DeepSeek V4.1 Flash
DeepSeek · open weights
98.97%$0.611337.2 sec
7Claude Sonnet 5
AnthropicEU data
98.63%$11.881783.8 sec
8GPT-6 Luna
OpenAIEU data
98.29%$0.661784.1 sec
9Gemini 3.8 Flash
Google
97.78%$6.65225.0 sec
10Mistral Large 4New
MistralEU data
97.09%$7.1011140.6 sec
11Qwen3.8 Flash
Qwen
96.92%$1.1513316.6 sec
12Mistral Medium 3.5
MistralEU data
96.92%$5.453332.8 sec
13Mistral Small 4
Mistral · open weightsEU data
92.99%$0.526442.4 sec
14Gemma 4 31B
Google · open weights
91.79%$0.194898.7 sec

Invoice extraction · German · Photo (15 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%75%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 100.00% · $0.93 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $1.11 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.45 / 1000Mistral Large 4 (Mistral) — 100.00% · $7.59 / 1000GPT-6 Sol (OpenAI) — 100.00% · $9.73 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $10.55 / 1000Claude Haiku 5.5 (Anthropic) — 99.49% · $0.68 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.49% · $0.70 / 1000Claude Sonnet 5 (Anthropic) — 98.97% · $11.05 / 1000GPT-6 Luna (OpenAI) — 98.46% · $0.50 / 1000Mistral Medium 3.5 (Mistral) — 94.87% · $5.68 / 1000Mistral Small 4 (Mistral) — 94.36% · $0.54 / 1000Gemma 4 31B (Google) — 93.33% · $0.19 / 1000Qwen3.8 Flash (Qwen) — 78.97% · $0.99 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5GLM-5.3 Flash100%95%90%85%80%75%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 100.00% · $0.93 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $1.11 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.45 / 1000Mistral Large 4 (Mistral) — 100.00% · $7.59 / 1000GPT-6 Sol (OpenAI) — 100.00% · $9.73 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $10.55 / 1000Claude Haiku 5.5 (Anthropic) — 99.49% · $0.68 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.49% · $0.70 / 1000Claude Sonnet 5 (Anthropic) — 98.97% · $11.05 / 1000GPT-6 Luna (OpenAI) — 98.46% · $0.50 / 1000Mistral Medium 3.5 (Mistral) — 94.87% · $5.68 / 1000Mistral Small 4 (Mistral) — 94.36% · $0.54 / 1000Gemma 4 31B (Google) — 93.33% · $0.19 / 1000Qwen3.8 Flash (Qwen) — 78.97% · $0.99 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5GLM-5.3 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.9304.4 sec
2Gemini 3.5 Flash Lite
GoogleEU data
100.00%$1.1101.6 sec
3Gemini 3.8 Flash
Google
100.00%$6.4504.8 sec
4Mistral Large 4New
MistralEU data
100.00%$7.59037.7 sec
5GPT-6 Sol
OpenAIEU data
100.00%$9.7305.4 sec
6Claude Sonnet 5.5New
AnthropicEU data
100.00%$10.5505.8 sec
7Claude Haiku 5.5New
AnthropicEU data
99.49%$0.68674.2 sec
8DeepSeek V4.1 Flash
DeepSeek · open weights
99.49%$0.70676.4 sec
9Claude Sonnet 5
AnthropicEU data
98.97%$11.051333.8 sec
10GPT-6 Luna
OpenAIEU data
98.46%$0.502003.5 sec
11Mistral Medium 3.5
MistralEU data
94.87%$5.684002.7 sec
12Mistral Small 4
Mistral · open weightsEU data
94.36%$0.544672.5 sec
13Gemma 4 31B
Google · open weights
93.33%$0.194679.1 sec
14Qwen3.8 Flash
Qwen
78.97%$0.9933315.6 sec

Invoice extraction · English (US) · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.61 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $9.54 / 1000GLM-5.3 Flash (Z.ai) — 99.94% · $0.95 / 1000GPT-6 Sol (OpenAI) — 99.94% · $9.65 / 1000Claude Sonnet 5 (Anthropic) — 99.82% · $8.84 / 1000Gemini 3.8 Flash (Google) — 99.82% · $3.84 / 1000GPT-6 Luna (OpenAI) — 99.70% · $0.51 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.39% · $0.52 / 1000Gemini 3.5 Flash Lite (Google) — 99.21% · $0.87 / 1000Mistral Medium 3.5 (Mistral) — 97.76% · $4.37 / 1000Gemma 4 31B (Google) — 97.64% · $0.17 / 1000Mistral Small 4 (Mistral) — 95.58% · $0.41 / 1000Mistral Large 4 (Mistral) — 95.39% · $6.65 / 1000Qwen3.8 Flash (Qwen) — 82.36% · $1.15 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5100%95%90%85%80%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.61 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $9.54 / 1000GLM-5.3 Flash (Z.ai) — 99.94% · $0.95 / 1000GPT-6 Sol (OpenAI) — 99.94% · $9.65 / 1000Claude Sonnet 5 (Anthropic) — 99.82% · $8.84 / 1000Gemini 3.8 Flash (Google) — 99.82% · $3.84 / 1000GPT-6 Luna (OpenAI) — 99.70% · $0.51 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.39% · $0.52 / 1000Gemini 3.5 Flash Lite (Google) — 99.21% · $0.87 / 1000Mistral Medium 3.5 (Mistral) — 97.76% · $4.37 / 1000Gemma 4 31B (Google) — 97.64% · $0.17 / 1000Mistral Small 4 (Mistral) — 95.58% · $0.41 / 1000Mistral Large 4 (Mistral) — 95.39% · $6.65 / 1000Qwen3.8 Flash (Qwen) — 82.36% · $1.15 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Haiku 5.5New
AnthropicEU data
100.00%$0.6103.4 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$9.5403.5 sec
3GLM-5.3 FlashNew
Z.ai · open weights
99.94%$0.9574.2 sec
4GPT-6 Sol
OpenAIEU data
99.94%$9.6574.1 sec
5Claude Sonnet 5
AnthropicEU data
99.82%$8.84133.2 sec
6Gemini 3.8 Flash
Google
99.82%$3.84203.2 sec
7GPT-6 Luna
OpenAIEU data
99.70%$0.51333.4 sec
8DeepSeek V4.1 Flash
DeepSeek · open weights
99.39%$0.52675.7 sec
9Gemini 3.5 Flash Lite
GoogleEU data
99.21%$0.87801.7 sec
10Mistral Medium 3.5
MistralEU data
97.76%$4.372072.3 sec
11Gemma 4 31B
Google · open weights
97.64%$0.171878.9 sec
12Mistral Small 4
Mistral · open weightsEU data
95.58%$0.414201.8 sec
13Mistral Large 4New
MistralEU data
95.39%$6.6522729.0 sec
14Qwen3.8 Flash
Qwen
82.36%$1.1542717.4 sec

Invoice extraction · English (US) · PDF text (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.35 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.52 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.53 / 1000Gemini 3.8 Flash (Google) — 100.00% · $2.92 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.28 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.86 / 1000Mistral Medium 3.5 (Mistral) — 99.80% · $2.24 / 1000Claude Sonnet 5 (Anthropic) — 99.80% · $3.91 / 1000Gemini 3.5 Flash Lite (Google) — 99.39% · $0.72 / 1000Mistral Small 4 (Mistral) — 98.38% · $0.19 / 1000Mistral Large 4 (Mistral) — 97.58% · $5.25 / 1000Qwen3.8 Flash (Qwen) — 91.72% · $0.97 / 1000GPT-6 Luna100%95%90%85%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.35 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.52 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.53 / 1000Gemini 3.8 Flash (Google) — 100.00% · $2.92 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.28 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.86 / 1000Mistral Medium 3.5 (Mistral) — 99.80% · $2.24 / 1000Claude Sonnet 5 (Anthropic) — 99.80% · $3.91 / 1000Gemini 3.5 Flash Lite (Google) — 99.39% · $0.72 / 1000Mistral Small 4 (Mistral) — 98.38% · $0.19 / 1000Mistral Large 4 (Mistral) — 97.58% · $5.25 / 1000Qwen3.8 Flash (Qwen) — 91.72% · $0.97 / 1000GPT-6 Luna
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GPT-6 Luna
OpenAIEU data
100.00%$0.1802.7 sec
2Gemma 4 31B
Google · open weights
100.00%$0.1807.4 sec
3Claude Haiku 5.5New
AnthropicEU data
100.00%$0.3502.4 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.5207.4 sec
5GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.5304.8 sec
6Gemini 3.8 Flash
Google
100.00%$2.9202.3 sec
7GPT-6 Sol
OpenAIEU data
100.00%$3.2803.2 sec
8Claude Sonnet 5.5New
AnthropicEU data
100.00%$4.8602.4 sec
9Mistral Medium 3.5
MistralEU data
99.80%$2.24221.2 sec
10Claude Sonnet 5
AnthropicEU data
99.80%$3.91222.4 sec
11Gemini 3.5 Flash Lite
GoogleEU data
99.39%$0.72441.0 sec
12Mistral Small 4
Mistral · open weightsEU data
98.38%$0.191561.2 sec
13Mistral Large 4New
MistralEU data
97.58%$5.254430.8 sec
14Qwen3.8 Flash
Qwen
91.72%$0.9751118.7 sec

Invoice extraction · English (US) · Clean image (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%75%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.75 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.23 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $12.19 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $12.58 / 1000GPT-6 Sol (OpenAI) — 100.00% · $13.81 / 1000GPT-6 Luna (OpenAI) — 99.60% · $0.73 / 1000Gemini 3.5 Flash Lite (Google) — 99.60% · $0.93 / 1000Gemini 3.8 Flash (Google) — 99.60% · $4.13 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.19% · $0.50 / 1000Gemma 4 31B (Google) — 98.18% · $0.16 / 1000Mistral Medium 3.5 (Mistral) — 97.98% · $5.25 / 1000Mistral Small 4 (Mistral) — 95.56% · $0.50 / 1000Mistral Large 4 (Mistral) — 95.35% · $6.57 / 1000Qwen3.8 Flash (Qwen) — 77.17% · $1.13 / 1000Gemma 4 31BDeepSeek V4.1 FlashGPT-6 LunaClaude Haiku 5.5100%95%90%85%80%75%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.75 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.23 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $12.19 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $12.58 / 1000GPT-6 Sol (OpenAI) — 100.00% · $13.81 / 1000GPT-6 Luna (OpenAI) — 99.60% · $0.73 / 1000Gemini 3.5 Flash Lite (Google) — 99.60% · $0.93 / 1000Gemini 3.8 Flash (Google) — 99.60% · $4.13 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.19% · $0.50 / 1000Gemma 4 31B (Google) — 98.18% · $0.16 / 1000Mistral Medium 3.5 (Mistral) — 97.98% · $5.25 / 1000Mistral Small 4 (Mistral) — 95.56% · $0.50 / 1000Mistral Large 4 (Mistral) — 95.35% · $6.57 / 1000Qwen3.8 Flash (Qwen) — 77.17% · $1.13 / 1000DeepSeek V4.1 FlashClaude Haiku 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Haiku 5.5New
AnthropicEU data
100.00%$0.7503.6 sec
2GLM-5.3 FlashNew
Z.ai · open weights
100.00%$1.2304.0 sec
3Claude Sonnet 5
AnthropicEU data
100.00%$12.1903.3 sec
4Claude Sonnet 5.5New
AnthropicEU data
100.00%$12.5803.6 sec
5GPT-6 Sol
OpenAIEU data
100.00%$13.8104.2 sec
6GPT-6 Luna
OpenAIEU data
99.60%$0.73443.8 sec
7Gemini 3.5 Flash Lite
GoogleEU data
99.60%$0.93441.8 sec
8Gemini 3.8 Flash
Google
99.60%$4.13443.5 sec
9DeepSeek V4.1 Flash
DeepSeek · open weights
99.19%$0.50894.6 sec
10Gemma 4 31B
Google · open weights
98.18%$0.161569.2 sec
11Mistral Medium 3.5
MistralEU data
97.98%$5.251782.3 sec
12Mistral Small 4
Mistral · open weightsEU data
95.56%$0.504441.7 sec
13Mistral Large 4New
MistralEU data
95.35%$6.5724424.2 sec
14Qwen3.8 Flash
Qwen
77.17%$1.1340015.2 sec

Invoice extraction · English (US) · Scan (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%75%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.71 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.08 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $10.66 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $11.43 / 1000GPT-6 Sol (OpenAI) — 100.00% · $12.04 / 1000Gemini 3.8 Flash (Google) — 99.80% · $4.36 / 1000GPT-6 Luna (OpenAI) — 99.60% · $0.63 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.99% · $0.55 / 1000Gemini 3.5 Flash Lite (Google) — 98.79% · $0.93 / 1000Gemma 4 31B (Google) — 96.57% · $0.16 / 1000Mistral Medium 3.5 (Mistral) — 95.96% · $5.27 / 1000Mistral Large 4 (Mistral) — 95.15% · $7.92 / 1000Mistral Small 4 (Mistral) — 93.54% · $0.50 / 1000Qwen3.8 Flash (Qwen) — 81.41% · $1.35 / 1000Gemma 4 31BDeepSeek V4.1 FlashGPT-6 LunaClaude Haiku 5.5100%95%90%85%80%75%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.71 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.08 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $10.66 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $11.43 / 1000GPT-6 Sol (OpenAI) — 100.00% · $12.04 / 1000Gemini 3.8 Flash (Google) — 99.80% · $4.36 / 1000GPT-6 Luna (OpenAI) — 99.60% · $0.63 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.99% · $0.55 / 1000Gemini 3.5 Flash Lite (Google) — 98.79% · $0.93 / 1000Gemma 4 31B (Google) — 96.57% · $0.16 / 1000Mistral Medium 3.5 (Mistral) — 95.96% · $5.27 / 1000Mistral Large 4 (Mistral) — 95.15% · $7.92 / 1000Mistral Small 4 (Mistral) — 93.54% · $0.50 / 1000Qwen3.8 Flash (Qwen) — 81.41% · $1.35 / 1000Gemma 4 31BClaude Haiku 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Haiku 5.5New
AnthropicEU data
100.00%$0.7103.9 sec
2GLM-5.3 FlashNew
Z.ai · open weights
100.00%$1.0804.0 sec
3Claude Sonnet 5
AnthropicEU data
100.00%$10.6603.3 sec
4Claude Sonnet 5.5New
AnthropicEU data
100.00%$11.4303.5 sec
5GPT-6 Sol
OpenAIEU data
100.00%$12.0404.3 sec
6Gemini 3.8 Flash
Google
99.80%$4.36223.2 sec
7GPT-6 Luna
OpenAIEU data
99.60%$0.63443.7 sec
8DeepSeek V4.1 Flash
DeepSeek · open weights
98.99%$0.551115.4 sec
9Gemini 3.5 Flash Lite
GoogleEU data
98.79%$0.931331.8 sec
10Gemma 4 31B
Google · open weights
96.57%$0.162679.8 sec
11Mistral Medium 3.5
MistralEU data
95.96%$5.273782.8 sec
12Mistral Large 4New
MistralEU data
95.15%$7.9231131.4 sec
13Mistral Small 4
Mistral · open weightsEU data
93.54%$0.506222.3 sec
14Qwen3.8 Flash
Qwen
81.41%$1.3540017.4 sec

Invoice extraction · English (US) · Photo (15 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%90%80%70%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.64 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.13 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $8.83 / 1000GPT-6 Luna (OpenAI) — 99.39% · $0.48 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.39% · $0.52 / 1000GLM-5.3 Flash (Z.ai) — 99.39% · $0.90 / 1000GPT-6 Sol (OpenAI) — 99.39% · $9.08 / 1000Claude Sonnet 5 (Anthropic) — 98.79% · $8.09 / 1000Gemini 3.5 Flash Lite (Google) — 98.79% · $0.94 / 1000Mistral Medium 3.5 (Mistral) — 96.36% · $5.44 / 1000Mistral Small 4 (Mistral) — 93.33% · $0.51 / 1000Gemma 4 31B (Google) — 92.12% · $0.16 / 1000Mistral Large 4 (Mistral) — 89.70% · $7.24 / 1000Qwen3.8 Flash (Qwen) — 72.73% · $1.20 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5100%90%80%70%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.64 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.13 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $8.83 / 1000GPT-6 Luna (OpenAI) — 99.39% · $0.48 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.39% · $0.52 / 1000GLM-5.3 Flash (Z.ai) — 99.39% · $0.90 / 1000GPT-6 Sol (OpenAI) — 99.39% · $9.08 / 1000Claude Sonnet 5 (Anthropic) — 98.79% · $8.09 / 1000Gemini 3.5 Flash Lite (Google) — 98.79% · $0.94 / 1000Mistral Medium 3.5 (Mistral) — 96.36% · $5.44 / 1000Mistral Small 4 (Mistral) — 93.33% · $0.51 / 1000Gemma 4 31B (Google) — 92.12% · $0.16 / 1000Mistral Large 4 (Mistral) — 89.70% · $7.24 / 1000Qwen3.8 Flash (Qwen) — 72.73% · $1.20 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Haiku 5.5New
AnthropicEU data
100.00%$0.6403.9 sec
2Gemini 3.8 Flash
Google
100.00%$4.1303.4 sec
3Claude Sonnet 5.5New
AnthropicEU data
100.00%$8.8303.4 sec
4GPT-6 Luna
OpenAIEU data
99.39%$0.48673.7 sec
5DeepSeek V4.1 Flash
DeepSeek · open weights
99.39%$0.52675.3 sec
6GLM-5.3 FlashNew
Z.ai · open weights
99.39%$0.90674.1 sec
7GPT-6 Sol
OpenAIEU data
99.39%$9.08674.9 sec
8Claude Sonnet 5
AnthropicEU data
98.79%$8.09673.2 sec
9Gemini 3.5 Flash Lite
GoogleEU data
98.79%$0.941331.7 sec
10Mistral Medium 3.5
MistralEU data
96.36%$5.443332.8 sec
11Mistral Small 4
Mistral · open weightsEU data
93.33%$0.515332.6 sec
12Gemma 4 31B
Google · open weights
92.12%$0.1660010.4 sec
13Mistral Large 4New
MistralEU data
89.70%$7.2446731.5 sec
14Qwen3.8 Flash
Qwen
72.73%$1.2033319.6 sec

Invoice extraction · Spanish · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $10.99 / 1000Claude Haiku 5.5 (Anthropic) — 99.44% · $0.73 / 1000GLM-5.3 Flash (Z.ai) — 99.38% · $1.06 / 1000GPT-6 Sol (OpenAI) — 99.23% · $10.79 / 1000GPT-6 Luna (OpenAI) — 99.13% · $0.55 / 1000Claude Sonnet 5 (Anthropic) — 98.46% · $11.51 / 1000Gemini 3.5 Flash Lite (Google) — 98.21% · $1.05 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.21% · $0.65 / 1000Gemini 3.8 Flash (Google) — 97.28% · $7.84 / 1000Mistral Medium 3.5 (Mistral) — 96.87% · $4.75 / 1000Mistral Large 4 (Mistral) — 96.51% · $7.49 / 1000Gemma 4 31B (Google) — 94.62% · $0.20 / 1000Mistral Small 4 (Mistral) — 92.36% · $0.44 / 1000Qwen3.8 Flash (Qwen) — 88.87% · $1.38 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5Claude Sonnet 5.5100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $10.99 / 1000Claude Haiku 5.5 (Anthropic) — 99.44% · $0.73 / 1000GLM-5.3 Flash (Z.ai) — 99.38% · $1.06 / 1000GPT-6 Sol (OpenAI) — 99.23% · $10.79 / 1000GPT-6 Luna (OpenAI) — 99.13% · $0.55 / 1000Claude Sonnet 5 (Anthropic) — 98.46% · $11.51 / 1000Gemini 3.5 Flash Lite (Google) — 98.21% · $1.05 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.21% · $0.65 / 1000Gemini 3.8 Flash (Google) — 97.28% · $7.84 / 1000Mistral Medium 3.5 (Mistral) — 96.87% · $4.75 / 1000Mistral Large 4 (Mistral) — 96.51% · $7.49 / 1000Gemma 4 31B (Google) — 94.62% · $0.20 / 1000Mistral Small 4 (Mistral) — 92.36% · $0.44 / 1000Qwen3.8 Flash (Qwen) — 88.87% · $1.38 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5Claude Sonnet 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5.5New
AnthropicEU data
100.00%$10.9905.7 sec
2Claude Haiku 5.5New
AnthropicEU data
99.44%$0.73734.4 sec
3GLM-5.3 FlashNew
Z.ai · open weights
99.38%$1.06735.0 sec
4GPT-6 Sol
OpenAIEU data
99.23%$10.791005.9 sec
5GPT-6 Luna
OpenAIEU data
99.13%$0.551133.9 sec
6Claude Sonnet 5
AnthropicEU data
98.46%$11.511874.2 sec
7Gemini 3.5 Flash Lite
GoogleEU data
98.21%$1.052201.6 sec
8DeepSeek V4.1 Flash
DeepSeek · open weights
98.21%$0.652276.3 sec
9Gemini 3.8 Flash
Google
97.28%$7.84335.7 sec
10Mistral Medium 3.5
MistralEU data
96.87%$4.753332.5 sec
11Mistral Large 4New
MistralEU data
96.51%$7.4912041.3 sec
12Gemma 4 31B
Google · open weights
94.62%$0.204937.4 sec
13Mistral Small 4
Mistral · open weightsEU data
92.36%$0.446732.1 sec
14Qwen3.8 Flash
Qwen
88.87%$1.3825322.0 sec

Invoice extraction · Spanish · PDF text (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 100.00% · $5.74 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.28 / 1000GPT-6 Luna (OpenAI) — 99.83% · $0.23 / 1000Qwen3.8 Flash (Qwen) — 99.83% · $1.03 / 1000GLM-5.3 Flash (Z.ai) — 99.66% · $0.65 / 1000GPT-6 Sol (OpenAI) — 99.66% · $4.26 / 1000Claude Haiku 5.5 (Anthropic) — 99.49% · $0.48 / 1000Mistral Large 4 (Mistral) — 99.49% · $5.90 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.97% · $0.55 / 1000Mistral Medium 3.5 (Mistral) — 98.97% · $3.01 / 1000Gemini 3.5 Flash Lite (Google) — 98.80% · $0.93 / 1000Claude Sonnet 5 (Anthropic) — 98.80% · $6.50 / 1000Gemma 4 31B (Google) — 98.46% · $0.23 / 1000Mistral Small 4 (Mistral) — 95.90% · $0.25 / 1000GPT-6 LunaGemini 3.8 Flash100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 100.00% · $5.74 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.28 / 1000GPT-6 Luna (OpenAI) — 99.83% · $0.23 / 1000Qwen3.8 Flash (Qwen) — 99.83% · $1.03 / 1000GLM-5.3 Flash (Z.ai) — 99.66% · $0.65 / 1000GPT-6 Sol (OpenAI) — 99.66% · $4.26 / 1000Claude Haiku 5.5 (Anthropic) — 99.49% · $0.48 / 1000Mistral Large 4 (Mistral) — 99.49% · $5.90 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.97% · $0.55 / 1000Mistral Medium 3.5 (Mistral) — 98.97% · $3.01 / 1000Gemini 3.5 Flash Lite (Google) — 98.80% · $0.93 / 1000Claude Sonnet 5 (Anthropic) — 98.80% · $6.50 / 1000Gemma 4 31B (Google) — 98.46% · $0.23 / 1000Mistral Small 4 (Mistral) — 95.90% · $0.25 / 1000GPT-6 LunaGemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.8 Flash
Google
100.00%$5.7404.7 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$6.2804.7 sec
3GPT-6 Luna
OpenAIEU data
99.83%$0.23223.0 sec
4Qwen3.8 Flash
Qwen
99.83%$1.032225.6 sec
5GLM-5.3 FlashNew
Z.ai · open weights
99.66%$0.65444.9 sec
6GPT-6 Sol
OpenAIEU data
99.66%$4.26444.5 sec
7Claude Haiku 5.5New
AnthropicEU data
99.49%$0.48673.2 sec
8Mistral Large 4New
MistralEU data
99.49%$5.906737.9 sec
9DeepSeek V4.1 Flash
DeepSeek · open weights
98.97%$0.551337.1 sec
10Mistral Medium 3.5
MistralEU data
98.97%$3.011331.4 sec
11Gemini 3.5 Flash Lite
GoogleEU data
98.80%$0.931561.1 sec
12Claude Sonnet 5
AnthropicEU data
98.80%$6.501564.9 sec
13Gemma 4 31B
Google · open weights
98.46%$0.232006.6 sec
14Mistral Small 4
Mistral · open weightsEU data
95.90%$0.254671.6 sec

Invoice extraction · Spanish · Clean image (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10$20Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $13.97 / 1000GLM-5.3 Flash (Z.ai) — 99.15% · $1.31 / 1000GPT-6 Luna (OpenAI) — 99.15% · $0.79 / 1000Claude Haiku 5.5 (Anthropic) — 98.97% · $0.90 / 1000GPT-6 Sol (OpenAI) — 98.97% · $15.49 / 1000Gemini 3.5 Flash Lite (Google) — 98.29% · $1.09 / 1000Claude Sonnet 5 (Anthropic) — 98.29% · $15.12 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.78% · $0.63 / 1000Mistral Large 4 (Mistral) — 96.92% · $7.79 / 1000Mistral Medium 3.5 (Mistral) — 96.92% · $5.44 / 1000Gemini 3.8 Flash (Google) — 95.56% · $8.64 / 1000Gemma 4 31B (Google) — 93.85% · $0.19 / 1000Mistral Small 4 (Mistral) — 90.77% · $0.52 / 1000Qwen3.8 Flash (Qwen) — 87.18% · $1.50 / 1000Gemma 4 31BDeepSeek V4.1 FlashGPT-6 LunaClaude Sonnet 5.5100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $13.97 / 1000GLM-5.3 Flash (Z.ai) — 99.15% · $1.31 / 1000GPT-6 Luna (OpenAI) — 99.15% · $0.79 / 1000Claude Haiku 5.5 (Anthropic) — 98.97% · $0.90 / 1000GPT-6 Sol (OpenAI) — 98.97% · $15.49 / 1000Gemini 3.5 Flash Lite (Google) — 98.29% · $1.09 / 1000Claude Sonnet 5 (Anthropic) — 98.29% · $15.12 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.78% · $0.63 / 1000Mistral Large 4 (Mistral) — 96.92% · $7.79 / 1000Mistral Medium 3.5 (Mistral) — 96.92% · $5.44 / 1000Gemini 3.8 Flash (Google) — 95.56% · $8.64 / 1000Gemma 4 31B (Google) — 93.85% · $0.19 / 1000Mistral Small 4 (Mistral) — 90.77% · $0.52 / 1000Qwen3.8 Flash (Qwen) — 87.18% · $1.50 / 1000Gemma 4 31BDeepSeek V4.1 FlashGPT-6 LunaClaude Sonnet 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5.5New
AnthropicEU data
100.00%$13.9705.3 sec
2GLM-5.3 FlashNew
Z.ai · open weights
99.15%$1.31895.0 sec
3GPT-6 Luna
OpenAIEU data
99.15%$0.791114.3 sec
4Claude Haiku 5.5New
AnthropicEU data
98.97%$0.901334.7 sec
5GPT-6 Sol
OpenAIEU data
98.97%$15.491336.7 sec
6Gemini 3.5 Flash Lite
GoogleEU data
98.29%$1.092221.8 sec
7Claude Sonnet 5
AnthropicEU data
98.29%$15.122224.4 sec
8DeepSeek V4.1 Flash
DeepSeek · open weights
97.78%$0.632896.3 sec
9Mistral Large 4New
MistralEU data
96.92%$7.7913339.3 sec
10Mistral Medium 3.5
MistralEU data
96.92%$5.443562.5 sec
11Gemini 3.8 Flash
Google
95.56%$8.64446.0 sec
12Gemma 4 31B
Google · open weights
93.85%$0.195787.9 sec
13Mistral Small 4
Mistral · open weightsEU data
90.77%$0.527562.0 sec
14Qwen3.8 Flash
Qwen
87.18%$1.5035620.0 sec

Invoice extraction · Spanish · Scan (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $12.89 / 1000Claude Haiku 5.5 (Anthropic) — 99.66% · $0.82 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $1.29 / 1000GPT-6 Sol (OpenAI) — 99.15% · $12.82 / 1000GPT-6 Luna (OpenAI) — 98.63% · $0.65 / 1000Claude Sonnet 5 (Anthropic) — 98.46% · $13.18 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.78% · $0.64 / 1000Gemini 3.5 Flash Lite (Google) — 97.44% · $1.10 / 1000Gemini 3.8 Flash (Google) — 95.38% · $9.38 / 1000Mistral Medium 3.5 (Mistral) — 95.38% · $5.47 / 1000Mistral Large 4 (Mistral) — 92.48% · $8.96 / 1000Gemma 4 31B (Google) — 91.97% · $0.19 / 1000Mistral Small 4 (Mistral) — 90.26% · $0.52 / 1000Qwen3.8 Flash (Qwen) — 83.25% · $1.58 / 1000Gemma 4 31BDeepSeek V4.1 FlashGPT-6 LunaClaude Haiku 5.5Claude Sonnet 5.5100%95%90%85%80%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $12.89 / 1000Claude Haiku 5.5 (Anthropic) — 99.66% · $0.82 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $1.29 / 1000GPT-6 Sol (OpenAI) — 99.15% · $12.82 / 1000GPT-6 Luna (OpenAI) — 98.63% · $0.65 / 1000Claude Sonnet 5 (Anthropic) — 98.46% · $13.18 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.78% · $0.64 / 1000Gemini 3.5 Flash Lite (Google) — 97.44% · $1.10 / 1000Gemini 3.8 Flash (Google) — 95.38% · $9.38 / 1000Mistral Medium 3.5 (Mistral) — 95.38% · $5.47 / 1000Mistral Large 4 (Mistral) — 92.48% · $8.96 / 1000Gemma 4 31B (Google) — 91.97% · $0.19 / 1000Mistral Small 4 (Mistral) — 90.26% · $0.52 / 1000Qwen3.8 Flash (Qwen) — 83.25% · $1.58 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5Claude Sonnet 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5.5New
AnthropicEU data
100.00%$12.8906.0 sec
2Claude Haiku 5.5New
AnthropicEU data
99.66%$0.82444.9 sec
3GLM-5.3 FlashNew
Z.ai · open weights
99.49%$1.29675.1 sec
4GPT-6 Sol
OpenAIEU data
99.15%$12.821116.1 sec
5GPT-6 Luna
OpenAIEU data
98.63%$0.651784.1 sec
6Claude Sonnet 5
AnthropicEU data
98.46%$13.181783.9 sec
7DeepSeek V4.1 Flash
DeepSeek · open weights
97.78%$0.642675.7 sec
8Gemini 3.5 Flash Lite
GoogleEU data
97.44%$1.102891.7 sec
9Gemini 3.8 Flash
Google
95.38%$9.38676.4 sec
10Mistral Medium 3.5
MistralEU data
95.38%$5.474672.7 sec
11Mistral Large 4New
MistralEU data
92.48%$8.9615646.6 sec
12Gemma 4 31B
Google · open weights
91.97%$0.196678.6 sec
13Mistral Small 4
Mistral · open weightsEU data
90.26%$0.528222.6 sec
14Qwen3.8 Flash
Qwen
83.25%$1.5831120.8 sec

Invoice extraction · Spanish · Photo (15 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%75%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.72 / 1000Gemini 3.8 Flash (Google) — 100.00% · $7.07 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $10.43 / 1000GLM-5.3 Flash (Z.ai) — 98.97% · $0.85 / 1000GPT-6 Sol (OpenAI) — 98.97% · $10.15 / 1000Mistral Large 4 (Mistral) — 98.46% · $6.93 / 1000GPT-6 Luna (OpenAI) — 98.46% · $0.50 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.46% · $1.04 / 1000Gemini 3.5 Flash Lite (Google) — 98.46% · $1.13 / 1000Claude Sonnet 5 (Anthropic) — 97.95% · $10.71 / 1000Mistral Medium 3.5 (Mistral) — 94.87% · $5.71 / 1000Gemma 4 31B (Google) — 93.33% · $0.19 / 1000Mistral Small 4 (Mistral) — 92.82% · $0.54 / 1000Qwen3.8 Flash (Qwen) — 77.95% · $1.51 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5100%95%90%85%80%75%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.72 / 1000Gemini 3.8 Flash (Google) — 100.00% · $7.07 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $10.43 / 1000GLM-5.3 Flash (Z.ai) — 98.97% · $0.85 / 1000GPT-6 Sol (OpenAI) — 98.97% · $10.15 / 1000Mistral Large 4 (Mistral) — 98.46% · $6.93 / 1000GPT-6 Luna (OpenAI) — 98.46% · $0.50 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.46% · $1.04 / 1000Gemini 3.5 Flash Lite (Google) — 98.46% · $1.13 / 1000Claude Sonnet 5 (Anthropic) — 97.95% · $10.71 / 1000Mistral Medium 3.5 (Mistral) — 94.87% · $5.71 / 1000Gemma 4 31B (Google) — 93.33% · $0.19 / 1000Mistral Small 4 (Mistral) — 92.82% · $0.54 / 1000Qwen3.8 Flash (Qwen) — 77.95% · $1.51 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Haiku 5.5New
AnthropicEU data
100.00%$0.7204.9 sec
2Gemini 3.8 Flash
Google
100.00%$7.0706.9 sec
3Claude Sonnet 5.5New
AnthropicEU data
100.00%$10.4306.4 sec
4GLM-5.3 FlashNew
Z.ai · open weights
98.97%$0.851334.7 sec
5GPT-6 Sol
OpenAIEU data
98.97%$10.151336.2 sec
6Mistral Large 4New
MistralEU data
98.46%$6.9313333.8 sec
7GPT-6 Luna
OpenAIEU data
98.46%$0.502004.4 sec
8DeepSeek V4.1 Flash
DeepSeek · open weights
98.46%$1.042007.3 sec
9Gemini 3.5 Flash Lite
GoogleEU data
98.46%$1.132001.7 sec
10Claude Sonnet 5
AnthropicEU data
97.95%$10.712004.0 sec
11Mistral Medium 3.5
MistralEU data
94.87%$5.714673.0 sec
12Gemma 4 31B
Google · open weights
93.33%$0.196007.4 sec
13Mistral Small 4
Mistral · open weightsEU data
92.82%$0.546002.8 sec
14Qwen3.8 Flash
Qwen
77.95%$1.5146718.3 sec

Invoice extraction · French · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 99.95% · $6.07 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $10.71 / 1000GPT-6 Sol (OpenAI) — 99.90% · $11.00 / 1000Claude Haiku 5.5 (Anthropic) — 99.85% · $0.77 / 1000Claude Sonnet 5 (Anthropic) — 99.69% · $10.69 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.54% · $0.60 / 1000GLM-5.3 Flash (Z.ai) — 99.08% · $0.98 / 1000GPT-6 Luna (OpenAI) — 99.08% · $0.56 / 1000Gemini 3.5 Flash Lite (Google) — 98.97% · $1.07 / 1000Mistral Medium 3.5 (Mistral) — 98.56% · $4.82 / 1000Mistral Small 4 (Mistral) — 94.87% · $0.44 / 1000Gemma 4 31B (Google) — 93.54% · $0.22 / 1000Qwen3.8 Flash (Qwen) — 90.46% · $1.28 / 1000Mistral Large 4 (Mistral) — 88.97% · $8.71 / 1000Gemma 4 31BMistral Small 4GPT-6 LunaDeepSeek V4.1 FlashClaude Haiku 5.5Gemini 3.8 Flash100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 99.95% · $6.07 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $10.71 / 1000GPT-6 Sol (OpenAI) — 99.90% · $11.00 / 1000Claude Haiku 5.5 (Anthropic) — 99.85% · $0.77 / 1000Claude Sonnet 5 (Anthropic) — 99.69% · $10.69 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.54% · $0.60 / 1000GLM-5.3 Flash (Z.ai) — 99.08% · $0.98 / 1000GPT-6 Luna (OpenAI) — 99.08% · $0.56 / 1000Gemini 3.5 Flash Lite (Google) — 98.97% · $1.07 / 1000Mistral Medium 3.5 (Mistral) — 98.56% · $4.82 / 1000Mistral Small 4 (Mistral) — 94.87% · $0.44 / 1000Gemma 4 31B (Google) — 93.54% · $0.22 / 1000Qwen3.8 Flash (Qwen) — 90.46% · $1.28 / 1000Mistral Large 4 (Mistral) — 88.97% · $8.71 / 1000Gemma 4 31BMistral Small 4GPT-6 LunaClaude Haiku 5.5Gemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.8 Flash
Google
99.95%$6.0775.3 sec
2Claude Sonnet 5.5New
AnthropicEU data
99.95%$10.7174.9 sec
3GPT-6 Sol
OpenAIEU data
99.90%$11.00136.0 sec
4Claude Haiku 5.5New
AnthropicEU data
99.85%$0.77204.6 sec
5Claude Sonnet 5
AnthropicEU data
99.69%$10.69334.4 sec
6DeepSeek V4.1 Flash
DeepSeek · open weights
99.54%$0.60606.2 sec
7GLM-5.3 FlashNew
Z.ai · open weights
99.08%$0.981074.7 sec
8GPT-6 Luna
OpenAIEU data
99.08%$0.561203.9 sec
9Gemini 3.5 Flash Lite
GoogleEU data
98.97%$1.071201.5 sec
10Mistral Medium 3.5
MistralEU data
98.56%$4.821532.6 sec
11Mistral Small 4
Mistral · open weightsEU data
94.87%$0.444732.1 sec
12Gemma 4 31B
Google · open weights
93.54%$0.224936.5 sec
13Qwen3.8 Flash
Qwen
90.46%$1.2816719.4 sec
14Mistral Large 4New
MistralEU data
88.97%$8.7134743.4 sec

Invoice extraction · French · PDF text (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.52 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.28 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.95 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.16 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.26 / 1000GPT-6 Luna (OpenAI) — 99.83% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.83% · $0.59 / 1000Qwen3.8 Flash (Qwen) — 99.83% · $1.06 / 1000Mistral Medium 3.5 (Mistral) — 99.83% · $3.13 / 1000Gemma 4 31B (Google) — 99.66% · $0.26 / 1000GLM-5.3 Flash (Z.ai) — 99.66% · $0.58 / 1000Gemini 3.5 Flash Lite (Google) — 99.49% · $0.96 / 1000Mistral Small 4 (Mistral) — 98.63% · $0.26 / 1000Mistral Large 4 (Mistral) — 97.44% · $6.23 / 1000GPT-6 LunaClaude Haiku 5.5100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.52 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.28 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.95 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.16 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.26 / 1000GPT-6 Luna (OpenAI) — 99.83% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.83% · $0.59 / 1000Qwen3.8 Flash (Qwen) — 99.83% · $1.06 / 1000Mistral Medium 3.5 (Mistral) — 99.83% · $3.13 / 1000Gemma 4 31B (Google) — 99.66% · $0.26 / 1000GLM-5.3 Flash (Z.ai) — 99.66% · $0.58 / 1000Gemini 3.5 Flash Lite (Google) — 99.49% · $0.96 / 1000Mistral Small 4 (Mistral) — 98.63% · $0.26 / 1000Mistral Large 4 (Mistral) — 97.44% · $6.23 / 1000Claude Haiku 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Haiku 5.5New
AnthropicEU data
100.00%$0.5203.4 sec
2GPT-6 Sol
OpenAIEU data
100.00%$4.2804.2 sec
3Gemini 3.8 Flash
Google
100.00%$4.9503.9 sec
4Claude Sonnet 5
AnthropicEU data
100.00%$6.1604.6 sec
5Claude Sonnet 5.5New
AnthropicEU data
100.00%$6.2603.0 sec
6GPT-6 Luna
OpenAIEU data
99.83%$0.23223.1 sec
7DeepSeek V4.1 Flash
DeepSeek · open weights
99.83%$0.59226.1 sec
8Qwen3.8 Flash
Qwen
99.83%$1.062224.4 sec
9Mistral Medium 3.5
MistralEU data
99.83%$3.13221.5 sec
10Gemma 4 31B
Google · open weights
99.66%$0.26445.9 sec
11GLM-5.3 FlashNew
Z.ai · open weights
99.66%$0.58445.0 sec
12Gemini 3.5 Flash Lite
GoogleEU data
99.49%$0.96671.1 sec
13Mistral Small 4
Mistral · open weightsEU data
98.63%$0.261561.4 sec
14Mistral Large 4New
MistralEU data
97.44%$6.236740.3 sec

Invoice extraction · French · Clean image (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%$0.20$0.50$1$2$5$10$20Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 100.00% · $6.33 / 1000GPT-6 Sol (OpenAI) — 100.00% · $15.44 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.93 / 1000Claude Sonnet 5.5 (Anthropic) — 99.83% · $13.55 / 1000Claude Sonnet 5 (Anthropic) — 99.66% · $13.17 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.49% · $0.58 / 1000Gemini 3.5 Flash Lite (Google) — 98.63% · $1.10 / 1000Mistral Medium 3.5 (Mistral) — 98.46% · $5.49 / 1000GPT-6 Luna (OpenAI) — 98.46% · $0.79 / 1000GLM-5.3 Flash (Z.ai) — 97.95% · $1.27 / 1000Mistral Small 4 (Mistral) — 93.33% · $0.52 / 1000Gemma 4 31B (Google) — 91.28% · $0.20 / 1000Mistral Large 4 (Mistral) — 87.86% · $9.95 / 1000Qwen3.8 Flash (Qwen) — 83.08% · $1.29 / 1000Gemma 4 31BMistral Small 4DeepSeek V4.1 FlashClaude Haiku 5.5Gemini 3.8 Flash100%95%90%85%80%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 100.00% · $6.33 / 1000GPT-6 Sol (OpenAI) — 100.00% · $15.44 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.93 / 1000Claude Sonnet 5.5 (Anthropic) — 99.83% · $13.55 / 1000Claude Sonnet 5 (Anthropic) — 99.66% · $13.17 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.49% · $0.58 / 1000Gemini 3.5 Flash Lite (Google) — 98.63% · $1.10 / 1000Mistral Medium 3.5 (Mistral) — 98.46% · $5.49 / 1000GPT-6 Luna (OpenAI) — 98.46% · $0.79 / 1000GLM-5.3 Flash (Z.ai) — 97.95% · $1.27 / 1000Mistral Small 4 (Mistral) — 93.33% · $0.52 / 1000Gemma 4 31B (Google) — 91.28% · $0.20 / 1000Mistral Large 4 (Mistral) — 87.86% · $9.95 / 1000Qwen3.8 Flash (Qwen) — 83.08% · $1.29 / 1000Gemma 4 31BMistral Small 4Claude Haiku 5.5Gemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.8 Flash
Google
100.00%$6.3305.5 sec
2GPT-6 Sol
OpenAIEU data
100.00%$15.4406.6 sec
3Claude Haiku 5.5New
AnthropicEU data
99.83%$0.93224.9 sec
4Claude Sonnet 5.5New
AnthropicEU data
99.83%$13.55225.1 sec
5Claude Sonnet 5
AnthropicEU data
99.66%$13.17444.3 sec
6DeepSeek V4.1 Flash
DeepSeek · open weights
99.49%$0.58676.1 sec
7Gemini 3.5 Flash Lite
GoogleEU data
98.63%$1.101781.7 sec
8Mistral Medium 3.5
MistralEU data
98.46%$5.491782.6 sec
9GPT-6 Luna
OpenAIEU data
98.46%$0.792004.3 sec
10GLM-5.3 FlashNew
Z.ai · open weights
97.95%$1.272224.4 sec
11Mistral Small 4
Mistral · open weightsEU data
93.33%$0.525782.0 sec
12Gemma 4 31B
Google · open weights
91.28%$0.206676.9 sec
13Mistral Large 4New
MistralEU data
87.86%$9.9546742.6 sec
14Qwen3.8 Flash
Qwen
83.08%$1.2922217.7 sec

Invoice extraction · French · Scan (45 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $12.55 / 1000GPT-6 Sol (OpenAI) — 100.00% · $13.54 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.87 / 1000Gemini 3.8 Flash (Google) — 99.83% · $6.56 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.66% · $0.62 / 1000Claude Sonnet 5 (Anthropic) — 99.49% · $12.49 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $1.13 / 1000GPT-6 Luna (OpenAI) — 98.97% · $0.68 / 1000Gemini 3.5 Flash Lite (Google) — 98.80% · $1.12 / 1000Mistral Medium 3.5 (Mistral) — 97.61% · $5.52 / 1000Mistral Small 4 (Mistral) — 93.16% · $0.52 / 1000Gemma 4 31B (Google) — 92.14% · $0.21 / 1000Qwen3.8 Flash (Qwen) — 88.21% · $1.48 / 1000Mistral Large 4 (Mistral) — 82.91% · $9.63 / 1000Gemma 4 31BMistral Small 4DeepSeek V4.1 FlashClaude Haiku 5.5Claude Sonnet 5.5100%95%90%85%80%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $12.55 / 1000GPT-6 Sol (OpenAI) — 100.00% · $13.54 / 1000Claude Haiku 5.5 (Anthropic) — 99.83% · $0.87 / 1000Gemini 3.8 Flash (Google) — 99.83% · $6.56 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.66% · $0.62 / 1000Claude Sonnet 5 (Anthropic) — 99.49% · $12.49 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $1.13 / 1000GPT-6 Luna (OpenAI) — 98.97% · $0.68 / 1000Gemini 3.5 Flash Lite (Google) — 98.80% · $1.12 / 1000Mistral Medium 3.5 (Mistral) — 97.61% · $5.52 / 1000Mistral Small 4 (Mistral) — 93.16% · $0.52 / 1000Gemma 4 31B (Google) — 92.14% · $0.21 / 1000Qwen3.8 Flash (Qwen) — 88.21% · $1.48 / 1000Mistral Large 4 (Mistral) — 82.91% · $9.63 / 1000Gemma 4 31BMistral Small 4Claude Haiku 5.5Claude Sonnet 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5.5New
AnthropicEU data
100.00%$12.5505.0 sec
2GPT-6 Sol
OpenAIEU data
100.00%$13.5406.2 sec
3Claude Haiku 5.5New
AnthropicEU data
99.83%$0.87224.9 sec
4Gemini 3.8 Flash
Google
99.83%$6.56225.7 sec
5DeepSeek V4.1 Flash
DeepSeek · open weights
99.66%$0.62446.1 sec
6Claude Sonnet 5
AnthropicEU data
99.49%$12.49444.4 sec
7GLM-5.3 FlashNew
Z.ai · open weights
99.49%$1.13674.4 sec
8GPT-6 Luna
OpenAIEU data
98.97%$0.681334.0 sec
9Gemini 3.5 Flash Lite
GoogleEU data
98.80%$1.121331.6 sec
10Mistral Medium 3.5
MistralEU data
97.61%$5.522672.9 sec
11Mistral Small 4
Mistral · open weightsEU data
93.16%$0.526222.6 sec
12Gemma 4 31B
Google · open weights
92.14%$0.216676.5 sec
13Qwen3.8 Flash
Qwen
88.21%$1.4820018.3 sec
14Mistral Large 4New
MistralEU data
82.91%$9.6351147.6 sec

Invoice extraction · French · Photo (15 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 100.00% · $7.13 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $9.97 / 1000Claude Haiku 5.5 (Anthropic) — 99.49% · $0.76 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $0.89 / 1000Claude Sonnet 5 (Anthropic) — 99.49% · $11.43 / 1000Gemini 3.5 Flash Lite (Google) — 98.97% · $1.14 / 1000GPT-6 Luna (OpenAI) — 98.97% · $0.50 / 1000GPT-6 Sol (OpenAI) — 98.97% · $10.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.46% · $0.61 / 1000Mistral Medium 3.5 (Mistral) — 97.95% · $5.76 / 1000Mistral Small 4 (Mistral) — 93.33% · $0.54 / 1000Qwen3.8 Flash (Qwen) — 91.28% · $1.32 / 1000Gemma 4 31B (Google) — 86.15% · $0.21 / 1000Mistral Large 4 (Mistral) — 85.13% · $9.74 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5Gemini 3.8 Flash100%95%90%85%80%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 100.00% · $7.13 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $9.97 / 1000Claude Haiku 5.5 (Anthropic) — 99.49% · $0.76 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $0.89 / 1000Claude Sonnet 5 (Anthropic) — 99.49% · $11.43 / 1000Gemini 3.5 Flash Lite (Google) — 98.97% · $1.14 / 1000GPT-6 Luna (OpenAI) — 98.97% · $0.50 / 1000GPT-6 Sol (OpenAI) — 98.97% · $10.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 98.46% · $0.61 / 1000Mistral Medium 3.5 (Mistral) — 97.95% · $5.76 / 1000Mistral Small 4 (Mistral) — 93.33% · $0.54 / 1000Qwen3.8 Flash (Qwen) — 91.28% · $1.32 / 1000Gemma 4 31B (Google) — 86.15% · $0.21 / 1000Mistral Large 4 (Mistral) — 85.13% · $9.74 / 1000Gemma 4 31BGPT-6 LunaClaude Haiku 5.5Gemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.8 Flash
Google
100.00%$7.1306.4 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$9.9704.6 sec
3Claude Haiku 5.5New
AnthropicEU data
99.49%$0.76674.9 sec
4GLM-5.3 FlashNew
Z.ai · open weights
99.49%$0.89674.0 sec
5Claude Sonnet 5
AnthropicEU data
99.49%$11.43675.4 sec
6Gemini 3.5 Flash Lite
GoogleEU data
98.97%$1.14671.6 sec
7GPT-6 Luna
OpenAIEU data
98.97%$0.501333.8 sec
8GPT-6 Sol
OpenAIEU data
98.97%$10.231336.9 sec
9DeepSeek V4.1 Flash
DeepSeek · open weights
98.46%$0.612007.7 sec
10Mistral Medium 3.5
MistralEU data
97.95%$5.761332.9 sec
11Mistral Small 4
Mistral · open weightsEU data
93.33%$0.546672.7 sec
12Qwen3.8 Flash
Qwen
91.28%$1.3233319.8 sec
13Gemma 4 31B
Google · open weights
86.15%$0.218007.2 sec
14Mistral Large 4New
MistralEU data
85.13%$9.7433349.7 sec

Email triage · Czech · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 99.19% · $3.06 / 1000GPT-6 Sol (OpenAI) — 98.96% · $3.19 / 1000GLM-5.3 Flash (Z.ai) — 98.89% · $0.52 / 1000Claude Sonnet 5.5 (Anthropic) — 98.67% · $3.92 / 1000Qwen3.8 Flash (Qwen) — 98.52% · $0.45 / 1000Mistral Large 4 (Mistral) — 98.15% · $4.41 / 1000Claude Sonnet 5 (Anthropic) — 98.15% · $4.91 / 1000GPT-6 Luna (OpenAI) — 97.93% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 97.63% · $0.29 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.41% · $0.30 / 1000Gemma 4 31B (Google) — 97.11% · $0.16 / 1000Gemini 3.5 Flash Lite (Google) — 95.78% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 94.52% · $1.95 / 1000Mistral Small 4 (Mistral) — 87.63% · $0.12 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaQwen3.8 FlashGLM-5.3 FlashGemini 3.8 Flash100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 99.19% · $3.06 / 1000GPT-6 Sol (OpenAI) — 98.96% · $3.19 / 1000GLM-5.3 Flash (Z.ai) — 98.89% · $0.52 / 1000Claude Sonnet 5.5 (Anthropic) — 98.67% · $3.92 / 1000Qwen3.8 Flash (Qwen) — 98.52% · $0.45 / 1000Mistral Large 4 (Mistral) — 98.15% · $4.41 / 1000Claude Sonnet 5 (Anthropic) — 98.15% · $4.91 / 1000GPT-6 Luna (OpenAI) — 97.93% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 97.63% · $0.29 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.41% · $0.30 / 1000Gemma 4 31B (Google) — 97.11% · $0.16 / 1000Gemini 3.5 Flash Lite (Google) — 95.78% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 94.52% · $1.95 / 1000Mistral Small 4 (Mistral) — 87.63% · $0.12 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaQwen3.8 FlashGLM-5.3 FlashGemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.8 Flash
Google
99.19%$3.06732.6 sec
2GPT-6 Sol
OpenAIEU data
98.96%$3.19932.8 sec
3GLM-5.3 FlashNew
Z.ai · open weights
98.89%$0.521004.7 sec
4Claude Sonnet 5.5New
AnthropicEU data
98.67%$3.921202.0 sec
5Qwen3.8 Flash
Qwen
98.52%$0.4513313.0 sec
6Mistral Large 4New
MistralEU data
98.15%$4.4110728.4 sec
7Claude Sonnet 5
AnthropicEU data
98.15%$4.911674.7 sec
8GPT-6 Luna
OpenAIEU data
97.93%$0.181872.6 sec
9Claude Haiku 5.5New
AnthropicEU data
97.63%$0.292001.9 sec
10DeepSeek V4.1 Flash
DeepSeek · open weights
97.41%$0.302274.2 sec
11Gemma 4 31B
Google · open weights
97.11%$0.162532.5 sec
12Gemini 3.5 Flash Lite
GoogleEU data
95.78%$0.513600.7 sec
13Mistral Medium 3.5
MistralEU data
94.52%$1.954470.8 sec
14Mistral Small 4
Mistral · open weightsEU data
87.63%$0.126870.8 sec

Email triage · Czech · Email (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 99.19% · $3.06 / 1000GPT-6 Sol (OpenAI) — 98.96% · $3.19 / 1000GLM-5.3 Flash (Z.ai) — 98.89% · $0.52 / 1000Claude Sonnet 5.5 (Anthropic) — 98.67% · $3.92 / 1000Qwen3.8 Flash (Qwen) — 98.52% · $0.45 / 1000Mistral Large 4 (Mistral) — 98.15% · $4.41 / 1000Claude Sonnet 5 (Anthropic) — 98.15% · $4.91 / 1000GPT-6 Luna (OpenAI) — 97.93% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 97.63% · $0.29 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.41% · $0.30 / 1000Gemma 4 31B (Google) — 97.11% · $0.16 / 1000Gemini 3.5 Flash Lite (Google) — 95.78% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 94.52% · $1.95 / 1000Mistral Small 4 (Mistral) — 87.63% · $0.12 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaQwen3.8 FlashGLM-5.3 FlashGemini 3.8 Flash100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 99.19% · $3.06 / 1000GPT-6 Sol (OpenAI) — 98.96% · $3.19 / 1000GLM-5.3 Flash (Z.ai) — 98.89% · $0.52 / 1000Claude Sonnet 5.5 (Anthropic) — 98.67% · $3.92 / 1000Qwen3.8 Flash (Qwen) — 98.52% · $0.45 / 1000Mistral Large 4 (Mistral) — 98.15% · $4.41 / 1000Claude Sonnet 5 (Anthropic) — 98.15% · $4.91 / 1000GPT-6 Luna (OpenAI) — 97.93% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 97.63% · $0.29 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.41% · $0.30 / 1000Gemma 4 31B (Google) — 97.11% · $0.16 / 1000Gemini 3.5 Flash Lite (Google) — 95.78% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 94.52% · $1.95 / 1000Mistral Small 4 (Mistral) — 87.63% · $0.12 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaQwen3.8 FlashGLM-5.3 FlashGemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.8 Flash
Google
99.19%$3.06732.6 sec
2GPT-6 Sol
OpenAIEU data
98.96%$3.19932.8 sec
3GLM-5.3 FlashNew
Z.ai · open weights
98.89%$0.521004.7 sec
4Claude Sonnet 5.5New
AnthropicEU data
98.67%$3.921202.0 sec
5Qwen3.8 Flash
Qwen
98.52%$0.4513313.0 sec
6Mistral Large 4New
MistralEU data
98.15%$4.4110728.4 sec
7Claude Sonnet 5
AnthropicEU data
98.15%$4.911674.7 sec
8GPT-6 Luna
OpenAIEU data
97.93%$0.181872.6 sec
9Claude Haiku 5.5New
AnthropicEU data
97.63%$0.292001.9 sec
10DeepSeek V4.1 Flash
DeepSeek · open weights
97.41%$0.302274.2 sec
11Gemma 4 31B
Google · open weights
97.11%$0.162532.5 sec
12Gemini 3.5 Flash Lite
GoogleEU data
95.78%$0.513600.7 sec
13Mistral Medium 3.5
MistralEU data
94.52%$1.954470.8 sec
14Mistral Small 4
Mistral · open weightsEU data
87.63%$0.126870.8 sec

Email triage · German · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 98.37% · $2.74 / 1000GPT-6 Sol (OpenAI) — 98.22% · $2.73 / 1000Claude Sonnet 5.5 (Anthropic) — 98.00% · $4.21 / 1000GLM-5.3 Flash (Z.ai) — 98.00% · $0.43 / 1000Mistral Large 4 (Mistral) — 97.85% · $3.78 / 1000Qwen3.8 Flash (Qwen) — 97.85% · $0.39 / 1000Claude Sonnet 5 (Anthropic) — 97.85% · $5.11 / 1000Claude Haiku 5.5 (Anthropic) — 97.11% · $0.32 / 1000GPT-6 Luna (OpenAI) — 96.96% · $0.15 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.37% · $0.30 / 1000Gemma 4 31B (Google) — 95.11% · $0.15 / 1000Gemini 3.5 Flash Lite (Google) — 94.67% · $0.48 / 1000Mistral Medium 3.5 (Mistral) — 92.74% · $1.77 / 1000Mistral Small 4 (Mistral) — 87.26% · $0.11 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaClaude Haiku 5.5Qwen3.8 FlashGLM-5.3 FlashGPT-6 SolGemini 3.8 Flash100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 98.37% · $2.74 / 1000GPT-6 Sol (OpenAI) — 98.22% · $2.73 / 1000Claude Sonnet 5.5 (Anthropic) — 98.00% · $4.21 / 1000GLM-5.3 Flash (Z.ai) — 98.00% · $0.43 / 1000Mistral Large 4 (Mistral) — 97.85% · $3.78 / 1000Qwen3.8 Flash (Qwen) — 97.85% · $0.39 / 1000Claude Sonnet 5 (Anthropic) — 97.85% · $5.11 / 1000Claude Haiku 5.5 (Anthropic) — 97.11% · $0.32 / 1000GPT-6 Luna (OpenAI) — 96.96% · $0.15 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.37% · $0.30 / 1000Gemma 4 31B (Google) — 95.11% · $0.15 / 1000Gemini 3.5 Flash Lite (Google) — 94.67% · $0.48 / 1000Mistral Medium 3.5 (Mistral) — 92.74% · $1.77 / 1000Mistral Small 4 (Mistral) — 87.26% · $0.11 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaClaude Haiku 5.5GLM-5.3 FlashGemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.8 Flash
Google
98.37%$2.741472.9 sec
2GPT-6 Sol
OpenAIEU data
98.22%$2.731602.9 sec
3Claude Sonnet 5.5New
AnthropicEU data
98.00%$4.211671.8 sec
4GLM-5.3 FlashNew
Z.ai · open weights
98.00%$0.431803.8 sec
5Mistral Large 4New
MistralEU data
97.85%$3.7818021.7 sec
6Qwen3.8 Flash
Qwen
97.85%$0.3918712.1 sec
7Claude Sonnet 5
AnthropicEU data
97.85%$5.111874.3 sec
8Claude Haiku 5.5New
AnthropicEU data
97.11%$0.322532.1 sec
9GPT-6 Luna
OpenAIEU data
96.96%$0.152732.8 sec
10DeepSeek V4.1 Flash
DeepSeek · open weights
96.37%$0.303204.2 sec
11Gemma 4 31B
Google · open weights
95.11%$0.153803.6 sec
12Gemini 3.5 Flash Lite
GoogleEU data
94.67%$0.484330.8 sec
13Mistral Medium 3.5
MistralEU data
92.74%$1.775200.8 sec
14Mistral Small 4
Mistral · open weightsEU data
87.26%$0.117070.8 sec

Email triage · German · Email (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 98.37% · $2.74 / 1000GPT-6 Sol (OpenAI) — 98.22% · $2.73 / 1000Claude Sonnet 5.5 (Anthropic) — 98.00% · $4.21 / 1000GLM-5.3 Flash (Z.ai) — 98.00% · $0.43 / 1000Mistral Large 4 (Mistral) — 97.85% · $3.78 / 1000Qwen3.8 Flash (Qwen) — 97.85% · $0.39 / 1000Claude Sonnet 5 (Anthropic) — 97.85% · $5.11 / 1000Claude Haiku 5.5 (Anthropic) — 97.11% · $0.32 / 1000GPT-6 Luna (OpenAI) — 96.96% · $0.15 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.37% · $0.30 / 1000Gemma 4 31B (Google) — 95.11% · $0.15 / 1000Gemini 3.5 Flash Lite (Google) — 94.67% · $0.48 / 1000Mistral Medium 3.5 (Mistral) — 92.74% · $1.77 / 1000Mistral Small 4 (Mistral) — 87.26% · $0.11 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaClaude Haiku 5.5Qwen3.8 FlashGLM-5.3 FlashGPT-6 SolGemini 3.8 Flash100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemini 3.8 Flash (Google) — 98.37% · $2.74 / 1000GPT-6 Sol (OpenAI) — 98.22% · $2.73 / 1000Claude Sonnet 5.5 (Anthropic) — 98.00% · $4.21 / 1000GLM-5.3 Flash (Z.ai) — 98.00% · $0.43 / 1000Mistral Large 4 (Mistral) — 97.85% · $3.78 / 1000Qwen3.8 Flash (Qwen) — 97.85% · $0.39 / 1000Claude Sonnet 5 (Anthropic) — 97.85% · $5.11 / 1000Claude Haiku 5.5 (Anthropic) — 97.11% · $0.32 / 1000GPT-6 Luna (OpenAI) — 96.96% · $0.15 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.37% · $0.30 / 1000Gemma 4 31B (Google) — 95.11% · $0.15 / 1000Gemini 3.5 Flash Lite (Google) — 94.67% · $0.48 / 1000Mistral Medium 3.5 (Mistral) — 92.74% · $1.77 / 1000Mistral Small 4 (Mistral) — 87.26% · $0.11 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaClaude Haiku 5.5GLM-5.3 FlashGemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemini 3.8 Flash
Google
98.37%$2.741472.9 sec
2GPT-6 Sol
OpenAIEU data
98.22%$2.731602.9 sec
3Claude Sonnet 5.5New
AnthropicEU data
98.00%$4.211671.8 sec
4GLM-5.3 FlashNew
Z.ai · open weights
98.00%$0.431803.8 sec
5Mistral Large 4New
MistralEU data
97.85%$3.7818021.7 sec
6Qwen3.8 Flash
Qwen
97.85%$0.3918712.1 sec
7Claude Sonnet 5
AnthropicEU data
97.85%$5.111874.3 sec
8Claude Haiku 5.5New
AnthropicEU data
97.11%$0.322532.1 sec
9GPT-6 Luna
OpenAIEU data
96.96%$0.152732.8 sec
10DeepSeek V4.1 Flash
DeepSeek · open weights
96.37%$0.303204.2 sec
11Gemma 4 31B
Google · open weights
95.11%$0.153803.6 sec
12Gemini 3.5 Flash Lite
GoogleEU data
94.67%$0.484330.8 sec
13Mistral Medium 3.5
MistralEU data
92.74%$1.775200.8 sec
14Mistral Small 4
Mistral · open weightsEU data
87.26%$0.117070.8 sec

Email triage · English (US) · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 99.41% · $0.33 / 1000GPT-6 Sol (OpenAI) — 99.04% · $2.44 / 1000Claude Sonnet 5.5 (Anthropic) — 98.89% · $3.21 / 1000Gemini 3.8 Flash (Google) — 98.74% · $2.39 / 1000Mistral Large 4 (Mistral) — 98.74% · $3.52 / 1000Claude Sonnet 5 (Anthropic) — 98.22% · $4.01 / 1000Claude Haiku 5.5 (Anthropic) — 98.15% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.63% · $0.23 / 1000Qwen3.8 Flash (Qwen) — 97.48% · $0.30 / 1000GPT-6 Luna (OpenAI) — 97.48% · $0.13 / 1000Gemma 4 31B (Google) — 96.52% · $0.12 / 1000Gemini 3.5 Flash Lite (Google) — 96.30% · $0.42 / 1000Mistral Medium 3.5 (Mistral) — 94.89% · $1.47 / 1000Mistral Small 4 (Mistral) — 89.70% · $0.09 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaClaude Haiku 5.5GLM-5.3 Flash100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 99.41% · $0.33 / 1000GPT-6 Sol (OpenAI) — 99.04% · $2.44 / 1000Claude Sonnet 5.5 (Anthropic) — 98.89% · $3.21 / 1000Gemini 3.8 Flash (Google) — 98.74% · $2.39 / 1000Mistral Large 4 (Mistral) — 98.74% · $3.52 / 1000Claude Sonnet 5 (Anthropic) — 98.22% · $4.01 / 1000Claude Haiku 5.5 (Anthropic) — 98.15% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.63% · $0.23 / 1000Qwen3.8 Flash (Qwen) — 97.48% · $0.30 / 1000GPT-6 Luna (OpenAI) — 97.48% · $0.13 / 1000Gemma 4 31B (Google) — 96.52% · $0.12 / 1000Gemini 3.5 Flash Lite (Google) — 96.30% · $0.42 / 1000Mistral Medium 3.5 (Mistral) — 94.89% · $1.47 / 1000Mistral Small 4 (Mistral) — 89.70% · $0.09 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaGLM-5.3 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GLM-5.3 FlashNew
Z.ai · open weights
99.41%$0.33533.9 sec
2GPT-6 Sol
OpenAIEU data
99.04%$2.44802.8 sec
3Claude Sonnet 5.5New
AnthropicEU data
98.89%$3.211002.0 sec
4Gemini 3.8 Flash
Google
98.74%$2.391132.2 sec
5Mistral Large 4New
MistralEU data
98.74%$3.5211319.8 sec
6Claude Sonnet 5
AnthropicEU data
98.22%$4.011603.8 sec
7Claude Haiku 5.5New
AnthropicEU data
98.15%$0.231601.7 sec
8DeepSeek V4.1 Flash
DeepSeek · open weights
97.63%$0.232133.8 sec
9Qwen3.8 Flash
Qwen
97.48%$0.302138.9 sec
10GPT-6 Luna
OpenAIEU data
97.48%$0.132272.5 sec
11Gemma 4 31B
Google · open weights
96.52%$0.122802.2 sec
12Gemini 3.5 Flash Lite
GoogleEU data
96.30%$0.423130.7 sec
13Mistral Medium 3.5
MistralEU data
94.89%$1.474270.8 sec
14Mistral Small 4
Mistral · open weightsEU data
89.70%$0.096270.8 sec

Email triage · English (US) · Email (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 99.41% · $0.33 / 1000GPT-6 Sol (OpenAI) — 99.04% · $2.44 / 1000Claude Sonnet 5.5 (Anthropic) — 98.89% · $3.21 / 1000Gemini 3.8 Flash (Google) — 98.74% · $2.39 / 1000Mistral Large 4 (Mistral) — 98.74% · $3.52 / 1000Claude Sonnet 5 (Anthropic) — 98.22% · $4.01 / 1000Claude Haiku 5.5 (Anthropic) — 98.15% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.63% · $0.23 / 1000Qwen3.8 Flash (Qwen) — 97.48% · $0.30 / 1000GPT-6 Luna (OpenAI) — 97.48% · $0.13 / 1000Gemma 4 31B (Google) — 96.52% · $0.12 / 1000Gemini 3.5 Flash Lite (Google) — 96.30% · $0.42 / 1000Mistral Medium 3.5 (Mistral) — 94.89% · $1.47 / 1000Mistral Small 4 (Mistral) — 89.70% · $0.09 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaClaude Haiku 5.5GLM-5.3 Flash100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 99.41% · $0.33 / 1000GPT-6 Sol (OpenAI) — 99.04% · $2.44 / 1000Claude Sonnet 5.5 (Anthropic) — 98.89% · $3.21 / 1000Gemini 3.8 Flash (Google) — 98.74% · $2.39 / 1000Mistral Large 4 (Mistral) — 98.74% · $3.52 / 1000Claude Sonnet 5 (Anthropic) — 98.22% · $4.01 / 1000Claude Haiku 5.5 (Anthropic) — 98.15% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.63% · $0.23 / 1000Qwen3.8 Flash (Qwen) — 97.48% · $0.30 / 1000GPT-6 Luna (OpenAI) — 97.48% · $0.13 / 1000Gemma 4 31B (Google) — 96.52% · $0.12 / 1000Gemini 3.5 Flash Lite (Google) — 96.30% · $0.42 / 1000Mistral Medium 3.5 (Mistral) — 94.89% · $1.47 / 1000Mistral Small 4 (Mistral) — 89.70% · $0.09 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaGLM-5.3 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GLM-5.3 FlashNew
Z.ai · open weights
99.41%$0.33533.9 sec
2GPT-6 Sol
OpenAIEU data
99.04%$2.44802.8 sec
3Claude Sonnet 5.5New
AnthropicEU data
98.89%$3.211002.0 sec
4Gemini 3.8 Flash
Google
98.74%$2.391132.2 sec
5Mistral Large 4New
MistralEU data
98.74%$3.5211319.8 sec
6Claude Sonnet 5
AnthropicEU data
98.22%$4.011603.8 sec
7Claude Haiku 5.5New
AnthropicEU data
98.15%$0.231601.7 sec
8DeepSeek V4.1 Flash
DeepSeek · open weights
97.63%$0.232133.8 sec
9Qwen3.8 Flash
Qwen
97.48%$0.302138.9 sec
10GPT-6 Luna
OpenAIEU data
97.48%$0.132272.5 sec
11Gemma 4 31B
Google · open weights
96.52%$0.122802.2 sec
12Gemini 3.5 Flash Lite
GoogleEU data
96.30%$0.423130.7 sec
13Mistral Medium 3.5
MistralEU data
94.89%$1.474270.8 sec
14Mistral Small 4
Mistral · open weightsEU data
89.70%$0.096270.8 sec

Email triage · Spanish · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 99.11% · $0.58 / 1000Gemini 3.8 Flash (Google) — 98.96% · $3.30 / 1000GPT-6 Sol (OpenAI) — 98.44% · $2.77 / 1000Mistral Large 4 (Mistral) — 97.93% · $4.03 / 1000Claude Sonnet 5.5 (Anthropic) — 97.70% · $3.76 / 1000Claude Haiku 5.5 (Anthropic) — 97.70% · $0.30 / 1000Claude Sonnet 5 (Anthropic) — 97.56% · $5.04 / 1000Gemma 4 31B (Google) — 97.11% · $0.15 / 1000GPT-6 Luna (OpenAI) — 96.81% · $0.16 / 1000Qwen3.8 Flash (Qwen) — 96.74% · $0.46 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.52% · $0.34 / 1000Gemini 3.5 Flash Lite (Google) — 95.56% · $0.48 / 1000Mistral Medium 3.5 (Mistral) — 94.15% · $1.79 / 1000Mistral Small 4 (Mistral) — 88.67% · $0.11 / 1000Mistral Small 4Gemma 4 31BClaude Haiku 5.5GLM-5.3 Flash100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 99.11% · $0.58 / 1000Gemini 3.8 Flash (Google) — 98.96% · $3.30 / 1000GPT-6 Sol (OpenAI) — 98.44% · $2.77 / 1000Mistral Large 4 (Mistral) — 97.93% · $4.03 / 1000Claude Sonnet 5.5 (Anthropic) — 97.70% · $3.76 / 1000Claude Haiku 5.5 (Anthropic) — 97.70% · $0.30 / 1000Claude Sonnet 5 (Anthropic) — 97.56% · $5.04 / 1000Gemma 4 31B (Google) — 97.11% · $0.15 / 1000GPT-6 Luna (OpenAI) — 96.81% · $0.16 / 1000Qwen3.8 Flash (Qwen) — 96.74% · $0.46 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.52% · $0.34 / 1000Gemini 3.5 Flash Lite (Google) — 95.56% · $0.48 / 1000Mistral Medium 3.5 (Mistral) — 94.15% · $1.79 / 1000Mistral Small 4 (Mistral) — 88.67% · $0.11 / 1000Mistral Small 4Gemma 4 31BClaude Haiku 5.5GLM-5.3 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GLM-5.3 FlashNew
Z.ai · open weights
99.11%$0.58804.5 sec
2Gemini 3.8 Flash
Google
98.96%$3.30933.3 sec
3GPT-6 Sol
OpenAIEU data
98.44%$2.771403.3 sec
4Mistral Large 4New
MistralEU data
97.93%$4.0318025.4 sec
5Claude Sonnet 5.5New
AnthropicEU data
97.70%$3.761931.8 sec
6Claude Haiku 5.5New
AnthropicEU data
97.70%$0.302002.1 sec
7Claude Sonnet 5
AnthropicEU data
97.56%$5.042135.5 sec
8Gemma 4 31B
Google · open weights
97.11%$0.152602.2 sec
9GPT-6 Luna
OpenAIEU data
96.81%$0.162803.1 sec
10Qwen3.8 Flash
Qwen
96.74%$0.4628011.2 sec
11DeepSeek V4.1 Flash
DeepSeek · open weights
96.52%$0.342805.5 sec
12Gemini 3.5 Flash Lite
GoogleEU data
95.56%$0.483470.8 sec
13Mistral Medium 3.5
MistralEU data
94.15%$1.794600.9 sec
14Mistral Small 4
Mistral · open weightsEU data
88.67%$0.116470.9 sec

Email triage · Spanish · Email (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 99.11% · $0.58 / 1000Gemini 3.8 Flash (Google) — 98.96% · $3.30 / 1000GPT-6 Sol (OpenAI) — 98.44% · $2.77 / 1000Mistral Large 4 (Mistral) — 97.93% · $4.03 / 1000Claude Sonnet 5.5 (Anthropic) — 97.70% · $3.76 / 1000Claude Haiku 5.5 (Anthropic) — 97.70% · $0.30 / 1000Claude Sonnet 5 (Anthropic) — 97.56% · $5.04 / 1000Gemma 4 31B (Google) — 97.11% · $0.15 / 1000GPT-6 Luna (OpenAI) — 96.81% · $0.16 / 1000Qwen3.8 Flash (Qwen) — 96.74% · $0.46 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.52% · $0.34 / 1000Gemini 3.5 Flash Lite (Google) — 95.56% · $0.48 / 1000Mistral Medium 3.5 (Mistral) — 94.15% · $1.79 / 1000Mistral Small 4 (Mistral) — 88.67% · $0.11 / 1000Mistral Small 4Gemma 4 31BClaude Haiku 5.5GLM-5.3 Flash100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 99.11% · $0.58 / 1000Gemini 3.8 Flash (Google) — 98.96% · $3.30 / 1000GPT-6 Sol (OpenAI) — 98.44% · $2.77 / 1000Mistral Large 4 (Mistral) — 97.93% · $4.03 / 1000Claude Sonnet 5.5 (Anthropic) — 97.70% · $3.76 / 1000Claude Haiku 5.5 (Anthropic) — 97.70% · $0.30 / 1000Claude Sonnet 5 (Anthropic) — 97.56% · $5.04 / 1000Gemma 4 31B (Google) — 97.11% · $0.15 / 1000GPT-6 Luna (OpenAI) — 96.81% · $0.16 / 1000Qwen3.8 Flash (Qwen) — 96.74% · $0.46 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.52% · $0.34 / 1000Gemini 3.5 Flash Lite (Google) — 95.56% · $0.48 / 1000Mistral Medium 3.5 (Mistral) — 94.15% · $1.79 / 1000Mistral Small 4 (Mistral) — 88.67% · $0.11 / 1000Mistral Small 4Gemma 4 31BClaude Haiku 5.5GLM-5.3 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GLM-5.3 FlashNew
Z.ai · open weights
99.11%$0.58804.5 sec
2Gemini 3.8 Flash
Google
98.96%$3.30933.3 sec
3GPT-6 Sol
OpenAIEU data
98.44%$2.771403.3 sec
4Mistral Large 4New
MistralEU data
97.93%$4.0318025.4 sec
5Claude Sonnet 5.5New
AnthropicEU data
97.70%$3.761931.8 sec
6Claude Haiku 5.5New
AnthropicEU data
97.70%$0.302002.1 sec
7Claude Sonnet 5
AnthropicEU data
97.56%$5.042135.5 sec
8Gemma 4 31B
Google · open weights
97.11%$0.152602.2 sec
9GPT-6 Luna
OpenAIEU data
96.81%$0.162803.1 sec
10Qwen3.8 Flash
Qwen
96.74%$0.4628011.2 sec
11DeepSeek V4.1 Flash
DeepSeek · open weights
96.52%$0.342805.5 sec
12Gemini 3.5 Flash Lite
GoogleEU data
95.56%$0.483470.8 sec
13Mistral Medium 3.5
MistralEU data
94.15%$1.794600.9 sec
14Mistral Small 4
Mistral · open weightsEU data
88.67%$0.116470.9 sec

Email triage · French · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5 (Anthropic) — 98.67% · $4.64 / 1000Gemini 3.8 Flash (Google) — 98.59% · $2.63 / 1000GPT-6 Sol (OpenAI) — 98.59% · $2.81 / 1000GLM-5.3 Flash (Z.ai) — 98.52% · $0.48 / 1000Mistral Large 4 (Mistral) — 98.30% · $3.70 / 1000Claude Sonnet 5.5 (Anthropic) — 97.78% · $3.85 / 1000Claude Haiku 5.5 (Anthropic) — 97.41% · $0.31 / 1000GPT-6 Luna (OpenAI) — 97.26% · $0.16 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.74% · $0.34 / 1000Qwen3.8 Flash (Qwen) — 96.44% · $0.45 / 1000Gemma 4 31B (Google) — 96.22% · $0.16 / 1000Gemini 3.5 Flash Lite (Google) — 96.15% · $0.49 / 1000Mistral Medium 3.5 (Mistral) — 94.15% · $1.78 / 1000Mistral Small 4 (Mistral) — 90.52% · $0.11 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaClaude Haiku 5.5GLM-5.3 FlashGemini 3.8 FlashClaude Sonnet 5100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5 (Anthropic) — 98.67% · $4.64 / 1000Gemini 3.8 Flash (Google) — 98.59% · $2.63 / 1000GPT-6 Sol (OpenAI) — 98.59% · $2.81 / 1000GLM-5.3 Flash (Z.ai) — 98.52% · $0.48 / 1000Mistral Large 4 (Mistral) — 98.30% · $3.70 / 1000Claude Sonnet 5.5 (Anthropic) — 97.78% · $3.85 / 1000Claude Haiku 5.5 (Anthropic) — 97.41% · $0.31 / 1000GPT-6 Luna (OpenAI) — 97.26% · $0.16 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.74% · $0.34 / 1000Qwen3.8 Flash (Qwen) — 96.44% · $0.45 / 1000Gemma 4 31B (Google) — 96.22% · $0.16 / 1000Gemini 3.5 Flash Lite (Google) — 96.15% · $0.49 / 1000Mistral Medium 3.5 (Mistral) — 94.15% · $1.78 / 1000Mistral Small 4 (Mistral) — 90.52% · $0.11 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaGLM-5.3 FlashGemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5
AnthropicEU data
98.67%$4.641204.3 sec
2Gemini 3.8 Flash
Google
98.59%$2.631272.7 sec
3GPT-6 Sol
OpenAIEU data
98.59%$2.811273.2 sec
4GLM-5.3 FlashNew
Z.ai · open weights
98.52%$0.481334.2 sec
5Mistral Large 4New
MistralEU data
98.30%$3.7015324.5 sec
6Claude Sonnet 5.5New
AnthropicEU data
97.78%$3.851801.7 sec
7Claude Haiku 5.5New
AnthropicEU data
97.41%$0.312332.2 sec
8GPT-6 Luna
OpenAIEU data
97.26%$0.162402.8 sec
9DeepSeek V4.1 Flash
DeepSeek · open weights
96.74%$0.342735.2 sec
10Qwen3.8 Flash
Qwen
96.44%$0.4524013.2 sec
11Gemma 4 31B
Google · open weights
96.22%$0.163132.4 sec
12Gemini 3.5 Flash Lite
GoogleEU data
96.15%$0.493130.8 sec
13Mistral Medium 3.5
MistralEU data
94.15%$1.784600.9 sec
14Mistral Small 4
Mistral · open weightsEU data
90.52%$0.115530.9 sec

Email triage · French · Email (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5 (Anthropic) — 98.67% · $4.64 / 1000Gemini 3.8 Flash (Google) — 98.59% · $2.63 / 1000GPT-6 Sol (OpenAI) — 98.59% · $2.81 / 1000GLM-5.3 Flash (Z.ai) — 98.52% · $0.48 / 1000Mistral Large 4 (Mistral) — 98.30% · $3.70 / 1000Claude Sonnet 5.5 (Anthropic) — 97.78% · $3.85 / 1000Claude Haiku 5.5 (Anthropic) — 97.41% · $0.31 / 1000GPT-6 Luna (OpenAI) — 97.26% · $0.16 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.74% · $0.34 / 1000Qwen3.8 Flash (Qwen) — 96.44% · $0.45 / 1000Gemma 4 31B (Google) — 96.22% · $0.16 / 1000Gemini 3.5 Flash Lite (Google) — 96.15% · $0.49 / 1000Mistral Medium 3.5 (Mistral) — 94.15% · $1.78 / 1000Mistral Small 4 (Mistral) — 90.52% · $0.11 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaClaude Haiku 5.5GLM-5.3 FlashGemini 3.8 FlashClaude Sonnet 5100%95%90%85%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5 (Anthropic) — 98.67% · $4.64 / 1000Gemini 3.8 Flash (Google) — 98.59% · $2.63 / 1000GPT-6 Sol (OpenAI) — 98.59% · $2.81 / 1000GLM-5.3 Flash (Z.ai) — 98.52% · $0.48 / 1000Mistral Large 4 (Mistral) — 98.30% · $3.70 / 1000Claude Sonnet 5.5 (Anthropic) — 97.78% · $3.85 / 1000Claude Haiku 5.5 (Anthropic) — 97.41% · $0.31 / 1000GPT-6 Luna (OpenAI) — 97.26% · $0.16 / 1000DeepSeek V4.1 Flash (DeepSeek) — 96.74% · $0.34 / 1000Qwen3.8 Flash (Qwen) — 96.44% · $0.45 / 1000Gemma 4 31B (Google) — 96.22% · $0.16 / 1000Gemini 3.5 Flash Lite (Google) — 96.15% · $0.49 / 1000Mistral Medium 3.5 (Mistral) — 94.15% · $1.78 / 1000Mistral Small 4 (Mistral) — 90.52% · $0.11 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaGLM-5.3 FlashGemini 3.8 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5
AnthropicEU data
98.67%$4.641204.3 sec
2Gemini 3.8 Flash
Google
98.59%$2.631272.7 sec
3GPT-6 Sol
OpenAIEU data
98.59%$2.811273.2 sec
4GLM-5.3 FlashNew
Z.ai · open weights
98.52%$0.481334.2 sec
5Mistral Large 4New
MistralEU data
98.30%$3.7015324.5 sec
6Claude Sonnet 5.5New
AnthropicEU data
97.78%$3.851801.7 sec
7Claude Haiku 5.5New
AnthropicEU data
97.41%$0.312332.2 sec
8GPT-6 Luna
OpenAIEU data
97.26%$0.162402.8 sec
9DeepSeek V4.1 Flash
DeepSeek · open weights
96.74%$0.342735.2 sec
10Qwen3.8 Flash
Qwen
96.44%$0.4524013.2 sec
11Gemma 4 31B
Google · open weights
96.22%$0.163132.4 sec
12Gemini 3.5 Flash Lite
GoogleEU data
96.15%$0.493130.8 sec
13Mistral Medium 3.5
MistralEU data
94.15%$1.784600.9 sec
14Mistral Small 4
Mistral · open weightsEU data
90.52%$0.115530.9 sec

Contract clauses · Czech · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.27 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.40 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.42 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.69 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $3.06 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.09 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.58 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.62 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.35 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.05 / 1000Gemma 4 31B (Google) — 99.95% · $0.25 / 1000Qwen3.8 Flash (Qwen) — 99.90% · $0.63 / 1000Gemini 3.5 Flash Lite (Google) — 99.85% · $0.85 / 1000Mistral Small 4 (Mistral) — 99.03% · $0.23 / 1000Mistral Small 4Gemma 4 31BGPT-6 Luna100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.27 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.40 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.42 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.69 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $3.06 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.09 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.58 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.62 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.35 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.05 / 1000Gemma 4 31B (Google) — 99.95% · $0.25 / 1000Qwen3.8 Flash (Qwen) — 99.90% · $0.63 / 1000Gemini 3.5 Flash Lite (Google) — 99.85% · $0.85 / 1000Mistral Small 4 (Mistral) — 99.03% · $0.23 / 1000Mistral Small 4Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GPT-6 Luna
OpenAIEU data
100.00%$0.2702.6 sec
2Claude Haiku 5.5New
AnthropicEU data
100.00%$0.4002.4 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.4204.3 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.6904.9 sec
5Mistral Medium 3.5
MistralEU data
100.00%$3.0601.2 sec
6Gemini 3.8 Flash
Google
100.00%$3.0902.2 sec
7Mistral Large 4New
MistralEU data
100.00%$4.58043.9 sec
8GPT-6 Sol
OpenAIEU data
100.00%$4.6202.0 sec
9Claude Sonnet 5.5New
AnthropicEU data
100.00%$5.3502.4 sec
10Claude Sonnet 5
AnthropicEU data
100.00%$6.0503.6 sec
11Gemma 4 31B
Google · open weights
99.95%$0.2576.0 sec
12Qwen3.8 Flash
Qwen
99.90%$0.631313.6 sec
13Gemini 3.5 Flash Lite
GoogleEU data
99.85%$0.85200.9 sec
14Mistral Small 4
Mistral · open weightsEU data
99.03%$0.23801.2 sec

Contract clauses · Czech · Contract text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.27 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.40 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.42 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.69 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $3.06 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.09 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.58 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.62 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.35 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.05 / 1000Gemma 4 31B (Google) — 99.95% · $0.25 / 1000Qwen3.8 Flash (Qwen) — 99.90% · $0.63 / 1000Gemini 3.5 Flash Lite (Google) — 99.85% · $0.85 / 1000Mistral Small 4 (Mistral) — 99.03% · $0.23 / 1000Mistral Small 4Gemma 4 31BGPT-6 Luna100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.27 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.40 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.42 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.69 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $3.06 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.09 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.58 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.62 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.35 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.05 / 1000Gemma 4 31B (Google) — 99.95% · $0.25 / 1000Qwen3.8 Flash (Qwen) — 99.90% · $0.63 / 1000Gemini 3.5 Flash Lite (Google) — 99.85% · $0.85 / 1000Mistral Small 4 (Mistral) — 99.03% · $0.23 / 1000Mistral Small 4Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GPT-6 Luna
OpenAIEU data
100.00%$0.2702.6 sec
2Claude Haiku 5.5New
AnthropicEU data
100.00%$0.4002.4 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.4204.3 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.6904.9 sec
5Mistral Medium 3.5
MistralEU data
100.00%$3.0601.2 sec
6Gemini 3.8 Flash
Google
100.00%$3.0902.2 sec
7Mistral Large 4New
MistralEU data
100.00%$4.58043.9 sec
8GPT-6 Sol
OpenAIEU data
100.00%$4.6202.0 sec
9Claude Sonnet 5.5New
AnthropicEU data
100.00%$5.3502.4 sec
10Claude Sonnet 5
AnthropicEU data
100.00%$6.0503.6 sec
11Gemma 4 31B
Google · open weights
99.95%$0.2576.0 sec
12Qwen3.8 Flash
Qwen
99.90%$0.631313.6 sec
13Gemini 3.5 Flash Lite
GoogleEU data
99.85%$0.85200.9 sec
14Mistral Small 4
Mistral · open weightsEU data
99.03%$0.23801.2 sec

Contract clauses · German · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.27 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.48 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.65 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.05 / 1000Gemini 3.8 Flash (Google) — 99.90% · $3.42 / 1000Mistral Large 4 (Mistral) — 99.44% · $5.12 / 1000Claude Sonnet 5 (Anthropic) — 98.21% · $7.11 / 1000Qwen3.8 Flash (Qwen) — 98.00% · $0.78 / 1000Gemini 3.5 Flash Lite (Google) — 98.00% · $0.79 / 1000Mistral Medium 3.5 (Mistral) — 97.13% · $2.76 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.08% · $0.52 / 1000Gemma 4 31B (Google) — 94.41% · $0.22 / 1000Mistral Small 4 (Mistral) — 94.26% · $0.21 / 1000Mistral Small 4Gemma 4 31BGPT-6 Luna100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.27 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.48 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.65 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.05 / 1000Gemini 3.8 Flash (Google) — 99.90% · $3.42 / 1000Mistral Large 4 (Mistral) — 99.44% · $5.12 / 1000Claude Sonnet 5 (Anthropic) — 98.21% · $7.11 / 1000Qwen3.8 Flash (Qwen) — 98.00% · $0.78 / 1000Gemini 3.5 Flash Lite (Google) — 98.00% · $0.79 / 1000Mistral Medium 3.5 (Mistral) — 97.13% · $2.76 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.08% · $0.52 / 1000Gemma 4 31B (Google) — 94.41% · $0.22 / 1000Mistral Small 4 (Mistral) — 94.26% · $0.21 / 1000Mistral Small 4Gemma 4 31BGPT-6 Luna
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GPT-6 Luna
OpenAIEU data
100.00%$0.2703.0 sec
2Claude Haiku 5.5New
AnthropicEU data
100.00%$0.4802.7 sec
3GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.6505.5 sec
4GPT-6 Sol
OpenAIEU data
100.00%$4.3602.9 sec
5Claude Sonnet 5.5New
AnthropicEU data
100.00%$6.0502.7 sec
6Gemini 3.8 Flash
Google
99.90%$3.4272.6 sec
7Mistral Large 4New
MistralEU data
99.44%$5.124736.2 sec
8Claude Sonnet 5
AnthropicEU data
98.21%$7.111334.3 sec
9Qwen3.8 Flash
Qwen
98.00%$0.7812719.7 sec
10Gemini 3.5 Flash Lite
GoogleEU data
98.00%$0.791601.0 sec
11Mistral Medium 3.5
MistralEU data
97.13%$2.762471.2 sec
12DeepSeek V4.1 Flash
DeepSeek · open weights
97.08%$0.522404.7 sec
13Gemma 4 31B
Google · open weights
94.41%$0.227272.6 sec
14Mistral Small 4
Mistral · open weightsEU data
94.26%$0.215331.2 sec

Contract clauses · German · Contract text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.27 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.48 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.65 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.05 / 1000Gemini 3.8 Flash (Google) — 99.90% · $3.42 / 1000Mistral Large 4 (Mistral) — 99.44% · $5.12 / 1000Claude Sonnet 5 (Anthropic) — 98.21% · $7.11 / 1000Qwen3.8 Flash (Qwen) — 98.00% · $0.78 / 1000Gemini 3.5 Flash Lite (Google) — 98.00% · $0.79 / 1000Mistral Medium 3.5 (Mistral) — 97.13% · $2.76 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.08% · $0.52 / 1000Gemma 4 31B (Google) — 94.41% · $0.22 / 1000Mistral Small 4 (Mistral) — 94.26% · $0.21 / 1000Mistral Small 4Gemma 4 31BGPT-6 Luna100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%GPT-6 Luna (OpenAI) — 100.00% · $0.27 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.48 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.65 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $6.05 / 1000Gemini 3.8 Flash (Google) — 99.90% · $3.42 / 1000Mistral Large 4 (Mistral) — 99.44% · $5.12 / 1000Claude Sonnet 5 (Anthropic) — 98.21% · $7.11 / 1000Qwen3.8 Flash (Qwen) — 98.00% · $0.78 / 1000Gemini 3.5 Flash Lite (Google) — 98.00% · $0.79 / 1000Mistral Medium 3.5 (Mistral) — 97.13% · $2.76 / 1000DeepSeek V4.1 Flash (DeepSeek) — 97.08% · $0.52 / 1000Gemma 4 31B (Google) — 94.41% · $0.22 / 1000Mistral Small 4 (Mistral) — 94.26% · $0.21 / 1000Mistral Small 4Gemma 4 31BGPT-6 Luna
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GPT-6 Luna
OpenAIEU data
100.00%$0.2703.0 sec
2Claude Haiku 5.5New
AnthropicEU data
100.00%$0.4802.7 sec
3GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.6505.5 sec
4GPT-6 Sol
OpenAIEU data
100.00%$4.3602.9 sec
5Claude Sonnet 5.5New
AnthropicEU data
100.00%$6.0502.7 sec
6Gemini 3.8 Flash
Google
99.90%$3.4272.6 sec
7Mistral Large 4New
MistralEU data
99.44%$5.124736.2 sec
8Claude Sonnet 5
AnthropicEU data
98.21%$7.111334.3 sec
9Qwen3.8 Flash
Qwen
98.00%$0.7812719.7 sec
10Gemini 3.5 Flash Lite
GoogleEU data
98.00%$0.791601.0 sec
11Mistral Medium 3.5
MistralEU data
97.13%$2.762471.2 sec
12DeepSeek V4.1 Flash
DeepSeek · open weights
97.08%$0.522404.7 sec
13Gemma 4 31B
Google · open weights
94.41%$0.227272.6 sec
14Mistral Small 4
Mistral · open weightsEU data
94.26%$0.215331.2 sec

Contract clauses · English (US) · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.32 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.33 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.04 / 1000Gemini 3.5 Flash Lite (Google) — 99.95% · $0.69 / 1000Gemini 3.8 Flash (Google) — 99.90% · $2.47 / 1000GPT-6 Luna (OpenAI) — 99.79% · $0.20 / 1000Claude Sonnet 5.5 (Anthropic) — 99.79% · $4.23 / 1000Claude Haiku 5.5 (Anthropic) — 99.69% · $0.35 / 1000Gemma 4 31B (Google) — 99.64% · $0.18 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $0.40 / 1000Mistral Medium 3.5 (Mistral) — 98.10% · $2.35 / 1000Mistral Large 4 (Mistral) — 96.10% · $5.24 / 1000Qwen3.8 Flash (Qwen) — 95.13% · $0.62 / 1000Mistral Small 4 (Mistral) — 91.54% · $0.17 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaDeepSeek V4.1 Flash100%95%90%85%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.32 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.33 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.04 / 1000Gemini 3.5 Flash Lite (Google) — 99.95% · $0.69 / 1000Gemini 3.8 Flash (Google) — 99.90% · $2.47 / 1000GPT-6 Luna (OpenAI) — 99.79% · $0.20 / 1000Claude Sonnet 5.5 (Anthropic) — 99.79% · $4.23 / 1000Claude Haiku 5.5 (Anthropic) — 99.69% · $0.35 / 1000Gemma 4 31B (Google) — 99.64% · $0.18 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $0.40 / 1000Mistral Medium 3.5 (Mistral) — 98.10% · $2.35 / 1000Mistral Large 4 (Mistral) — 96.10% · $5.24 / 1000Qwen3.8 Flash (Qwen) — 95.13% · $0.62 / 1000Mistral Small 4 (Mistral) — 91.54% · $0.17 / 1000Mistral Small 4Gemma 4 31BDeepSeek V4.1 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3203.9 sec
2GPT-6 Sol
OpenAIEU data
100.00%$3.3302.5 sec
3Claude Sonnet 5
AnthropicEU data
100.00%$5.0403.4 sec
4Gemini 3.5 Flash Lite
GoogleEU data
99.95%$0.6971.0 sec
5Gemini 3.8 Flash
Google
99.90%$2.4772.0 sec
6GPT-6 Luna
OpenAIEU data
99.79%$0.20132.7 sec
7Claude Sonnet 5.5New
AnthropicEU data
99.79%$4.23202.1 sec
8Claude Haiku 5.5New
AnthropicEU data
99.69%$0.35272.1 sec
9Gemma 4 31B
Google · open weights
99.64%$0.18272.5 sec
10GLM-5.3 FlashNew
Z.ai · open weights
99.49%$0.40403.4 sec
11Mistral Medium 3.5
MistralEU data
98.10%$2.351271.1 sec
12Mistral Large 4New
MistralEU data
96.10%$5.2425341.4 sec
13Qwen3.8 Flash
Qwen
95.13%$0.6230714.1 sec
14Mistral Small 4
Mistral · open weightsEU data
91.54%$0.175671.1 sec

Contract clauses · English (US) · Contract text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.32 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.33 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.04 / 1000Gemini 3.5 Flash Lite (Google) — 99.95% · $0.69 / 1000Gemini 3.8 Flash (Google) — 99.90% · $2.47 / 1000GPT-6 Luna (OpenAI) — 99.79% · $0.20 / 1000Claude Sonnet 5.5 (Anthropic) — 99.79% · $4.23 / 1000Claude Haiku 5.5 (Anthropic) — 99.69% · $0.35 / 1000Gemma 4 31B (Google) — 99.64% · $0.18 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $0.40 / 1000Mistral Medium 3.5 (Mistral) — 98.10% · $2.35 / 1000Mistral Large 4 (Mistral) — 96.10% · $5.24 / 1000Qwen3.8 Flash (Qwen) — 95.13% · $0.62 / 1000Mistral Small 4 (Mistral) — 91.54% · $0.17 / 1000Mistral Small 4Gemma 4 31BGPT-6 LunaDeepSeek V4.1 Flash100%95%90%85%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.32 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.33 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.04 / 1000Gemini 3.5 Flash Lite (Google) — 99.95% · $0.69 / 1000Gemini 3.8 Flash (Google) — 99.90% · $2.47 / 1000GPT-6 Luna (OpenAI) — 99.79% · $0.20 / 1000Claude Sonnet 5.5 (Anthropic) — 99.79% · $4.23 / 1000Claude Haiku 5.5 (Anthropic) — 99.69% · $0.35 / 1000Gemma 4 31B (Google) — 99.64% · $0.18 / 1000GLM-5.3 Flash (Z.ai) — 99.49% · $0.40 / 1000Mistral Medium 3.5 (Mistral) — 98.10% · $2.35 / 1000Mistral Large 4 (Mistral) — 96.10% · $5.24 / 1000Qwen3.8 Flash (Qwen) — 95.13% · $0.62 / 1000Mistral Small 4 (Mistral) — 91.54% · $0.17 / 1000Mistral Small 4Gemma 4 31BDeepSeek V4.1 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3203.9 sec
2GPT-6 Sol
OpenAIEU data
100.00%$3.3302.5 sec
3Claude Sonnet 5
AnthropicEU data
100.00%$5.0403.4 sec
4Gemini 3.5 Flash Lite
GoogleEU data
99.95%$0.6971.0 sec
5Gemini 3.8 Flash
Google
99.90%$2.4772.0 sec
6GPT-6 Luna
OpenAIEU data
99.79%$0.20132.7 sec
7Claude Sonnet 5.5New
AnthropicEU data
99.79%$4.23202.1 sec
8Claude Haiku 5.5New
AnthropicEU data
99.69%$0.35272.1 sec
9Gemma 4 31B
Google · open weights
99.64%$0.18272.5 sec
10GLM-5.3 FlashNew
Z.ai · open weights
99.49%$0.40403.4 sec
11Mistral Medium 3.5
MistralEU data
98.10%$2.351271.1 sec
12Mistral Large 4New
MistralEU data
96.10%$5.2425341.4 sec
13Qwen3.8 Flash
Qwen
95.13%$0.6230714.1 sec
14Mistral Small 4
Mistral · open weightsEU data
91.54%$0.175671.1 sec

Contract clauses · Spanish · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.23 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.26 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.40 / 1000Qwen3.8 Flash (Qwen) — 100.00% · $0.60 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.64 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $3.02 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.33 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.48 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.83 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.70 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.58 / 1000Claude Haiku 5.5 (Anthropic) — 99.95% · $0.42 / 1000Gemini 3.5 Flash Lite (Google) — 99.79% · $0.84 / 1000Gemma 4 31B (Google) — 94.41% · $0.24 / 1000Mistral Small 4100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.23 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.26 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.40 / 1000Qwen3.8 Flash (Qwen) — 100.00% · $0.60 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.64 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $3.02 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.33 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.48 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.83 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.70 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.58 / 1000Claude Haiku 5.5 (Anthropic) — 99.95% · $0.42 / 1000Gemini 3.5 Flash Lite (Google) — 99.79% · $0.84 / 1000Gemma 4 31B (Google) — 94.41% · $0.24 / 1000Mistral Small 4
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Small 4
Mistral · open weightsEU data
100.00%$0.2301.3 sec
2GPT-6 Luna
OpenAIEU data
100.00%$0.2603.1 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.4004.1 sec
4Qwen3.8 Flash
Qwen
100.00%$0.60014.2 sec
5GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.6404.9 sec
6Mistral Medium 3.5
MistralEU data
100.00%$3.0201.3 sec
7Gemini 3.8 Flash
Google
100.00%$3.3303.0 sec
8GPT-6 Sol
OpenAIEU data
100.00%$4.4802.8 sec
9Mistral Large 4New
MistralEU data
100.00%$4.83035.0 sec
10Claude Sonnet 5.5New
AnthropicEU data
100.00%$5.7002.7 sec
11Claude Sonnet 5
AnthropicEU data
100.00%$6.5804.0 sec
12Claude Haiku 5.5New
AnthropicEU data
99.95%$0.4272.3 sec
13Gemini 3.5 Flash Lite
GoogleEU data
99.79%$0.84201.1 sec
14Gemma 4 31B
Google · open weights
94.41%$0.247275.0 sec

Contract clauses · Spanish · Contract text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.23 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.26 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.40 / 1000Qwen3.8 Flash (Qwen) — 100.00% · $0.60 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.64 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $3.02 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.33 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.48 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.83 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.70 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.58 / 1000Claude Haiku 5.5 (Anthropic) — 99.95% · $0.42 / 1000Gemini 3.5 Flash Lite (Google) — 99.79% · $0.84 / 1000Gemma 4 31B (Google) — 94.41% · $0.24 / 1000Mistral Small 4100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.23 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.26 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.40 / 1000Qwen3.8 Flash (Qwen) — 100.00% · $0.60 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.64 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $3.02 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.33 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.48 / 1000Mistral Large 4 (Mistral) — 100.00% · $4.83 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.70 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $6.58 / 1000Claude Haiku 5.5 (Anthropic) — 99.95% · $0.42 / 1000Gemini 3.5 Flash Lite (Google) — 99.79% · $0.84 / 1000Gemma 4 31B (Google) — 94.41% · $0.24 / 1000Mistral Small 4
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Small 4
Mistral · open weightsEU data
100.00%$0.2301.3 sec
2GPT-6 Luna
OpenAIEU data
100.00%$0.2603.1 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.4004.1 sec
4Qwen3.8 Flash
Qwen
100.00%$0.60014.2 sec
5GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.6404.9 sec
6Mistral Medium 3.5
MistralEU data
100.00%$3.0201.3 sec
7Gemini 3.8 Flash
Google
100.00%$3.3303.0 sec
8GPT-6 Sol
OpenAIEU data
100.00%$4.4802.8 sec
9Mistral Large 4New
MistralEU data
100.00%$4.83035.0 sec
10Claude Sonnet 5.5New
AnthropicEU data
100.00%$5.7002.7 sec
11Claude Sonnet 5
AnthropicEU data
100.00%$6.5804.0 sec
12Claude Haiku 5.5New
AnthropicEU data
99.95%$0.4272.3 sec
13Gemini 3.5 Flash Lite
GoogleEU data
99.79%$0.84201.1 sec
14Gemma 4 31B
Google · open weights
94.41%$0.247275.0 sec

Contract clauses · French · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Medium 3.5 (Mistral) — 100.00% · $3.01 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.50 / 1000Gemini 3.5 Flash Lite (Google) — 99.95% · $0.85 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $5.71 / 1000Claude Sonnet 5 (Anthropic) — 99.95% · $6.69 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.85% · $0.45 / 1000GLM-5.3 Flash (Z.ai) — 99.74% · $0.69 / 1000GPT-6 Sol (OpenAI) — 99.59% · $4.49 / 1000Qwen3.8 Flash (Qwen) — 99.49% · $0.66 / 1000Claude Haiku 5.5 (Anthropic) — 99.38% · $0.45 / 1000GPT-6 Luna (OpenAI) — 99.33% · $0.27 / 1000Mistral Large 4 (Mistral) — 99.28% · $4.60 / 1000Mistral Small 4 (Mistral) — 99.23% · $0.23 / 1000Gemma 4 31B (Google) — 94.62% · $0.26 / 1000Mistral Small 4GPT-6 LunaDeepSeek V4.1 FlashGemini 3.5 Flash LiteMistral Medium 3.5100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Medium 3.5 (Mistral) — 100.00% · $3.01 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.50 / 1000Gemini 3.5 Flash Lite (Google) — 99.95% · $0.85 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $5.71 / 1000Claude Sonnet 5 (Anthropic) — 99.95% · $6.69 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.85% · $0.45 / 1000GLM-5.3 Flash (Z.ai) — 99.74% · $0.69 / 1000GPT-6 Sol (OpenAI) — 99.59% · $4.49 / 1000Qwen3.8 Flash (Qwen) — 99.49% · $0.66 / 1000Claude Haiku 5.5 (Anthropic) — 99.38% · $0.45 / 1000GPT-6 Luna (OpenAI) — 99.33% · $0.27 / 1000Mistral Large 4 (Mistral) — 99.28% · $4.60 / 1000Mistral Small 4 (Mistral) — 99.23% · $0.23 / 1000Gemma 4 31B (Google) — 94.62% · $0.26 / 1000Mistral Small 4GPT-6 LunaGemini 3.5 Flash Lite
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Medium 3.5
MistralEU data
100.00%$3.0101.3 sec
2Gemini 3.8 Flash
Google
100.00%$3.5003.0 sec
3Gemini 3.5 Flash Lite
GoogleEU data
99.95%$0.8571.0 sec
4Claude Sonnet 5.5New
AnthropicEU data
99.95%$5.7172.6 sec
5Claude Sonnet 5
AnthropicEU data
99.95%$6.6974.6 sec
6DeepSeek V4.1 Flash
DeepSeek · open weights
99.85%$0.45205.2 sec
7GLM-5.3 FlashNew
Z.ai · open weights
99.74%$0.69335.1 sec
8GPT-6 Sol
OpenAIEU data
99.59%$4.49532.6 sec
9Qwen3.8 Flash
Qwen
99.49%$0.666716.2 sec
10Claude Haiku 5.5New
AnthropicEU data
99.38%$0.45802.6 sec
11GPT-6 Luna
OpenAIEU data
99.33%$0.27873.2 sec
12Mistral Large 4New
MistralEU data
99.28%$4.609334.7 sec
13Mistral Small 4
Mistral · open weightsEU data
99.23%$0.231001.2 sec
14Gemma 4 31B
Google · open weights
94.62%$0.266935.3 sec

Contract clauses · French · Contract text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Medium 3.5 (Mistral) — 100.00% · $3.01 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.50 / 1000Gemini 3.5 Flash Lite (Google) — 99.95% · $0.85 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $5.71 / 1000Claude Sonnet 5 (Anthropic) — 99.95% · $6.69 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.85% · $0.45 / 1000GLM-5.3 Flash (Z.ai) — 99.74% · $0.69 / 1000GPT-6 Sol (OpenAI) — 99.59% · $4.49 / 1000Qwen3.8 Flash (Qwen) — 99.49% · $0.66 / 1000Claude Haiku 5.5 (Anthropic) — 99.38% · $0.45 / 1000GPT-6 Luna (OpenAI) — 99.33% · $0.27 / 1000Mistral Large 4 (Mistral) — 99.28% · $4.60 / 1000Mistral Small 4 (Mistral) — 99.23% · $0.23 / 1000Gemma 4 31B (Google) — 94.62% · $0.26 / 1000Mistral Small 4GPT-6 LunaDeepSeek V4.1 FlashGemini 3.5 Flash LiteMistral Medium 3.5100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Medium 3.5 (Mistral) — 100.00% · $3.01 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.50 / 1000Gemini 3.5 Flash Lite (Google) — 99.95% · $0.85 / 1000Claude Sonnet 5.5 (Anthropic) — 99.95% · $5.71 / 1000Claude Sonnet 5 (Anthropic) — 99.95% · $6.69 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.85% · $0.45 / 1000GLM-5.3 Flash (Z.ai) — 99.74% · $0.69 / 1000GPT-6 Sol (OpenAI) — 99.59% · $4.49 / 1000Qwen3.8 Flash (Qwen) — 99.49% · $0.66 / 1000Claude Haiku 5.5 (Anthropic) — 99.38% · $0.45 / 1000GPT-6 Luna (OpenAI) — 99.33% · $0.27 / 1000Mistral Large 4 (Mistral) — 99.28% · $4.60 / 1000Mistral Small 4 (Mistral) — 99.23% · $0.23 / 1000Gemma 4 31B (Google) — 94.62% · $0.26 / 1000Mistral Small 4GPT-6 LunaGemini 3.5 Flash Lite
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Medium 3.5
MistralEU data
100.00%$3.0101.3 sec
2Gemini 3.8 Flash
Google
100.00%$3.5003.0 sec
3Gemini 3.5 Flash Lite
GoogleEU data
99.95%$0.8571.0 sec
4Claude Sonnet 5.5New
AnthropicEU data
99.95%$5.7172.6 sec
5Claude Sonnet 5
AnthropicEU data
99.95%$6.6974.6 sec
6DeepSeek V4.1 Flash
DeepSeek · open weights
99.85%$0.45205.2 sec
7GLM-5.3 FlashNew
Z.ai · open weights
99.74%$0.69335.1 sec
8GPT-6 Sol
OpenAIEU data
99.59%$4.49532.6 sec
9Qwen3.8 Flash
Qwen
99.49%$0.666716.2 sec
10Claude Haiku 5.5New
AnthropicEU data
99.38%$0.45802.6 sec
11GPT-6 Luna
OpenAIEU data
99.33%$0.27873.2 sec
12Mistral Large 4New
MistralEU data
99.28%$4.609334.7 sec
13Mistral Small 4
Mistral · open weightsEU data
99.23%$0.231001.2 sec
14Gemma 4 31B
Google · open weights
94.62%$0.266935.3 sec

Personal data detection · Czech · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 99.81% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.81% · $1.40 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.71% · $0.31 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.22 / 1000Claude Sonnet 5.5 (Anthropic) — 99.71% · $3.52 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Gemini 3.8 Flash (Google) — 99.52% · $2.19 / 1000Claude Haiku 5.5 (Anthropic) — 99.43% · $0.21 / 1000GLM-5.3 Flash (Z.ai) — 99.33% · $0.49 / 1000Mistral Large 4 (Mistral) — 99.33% · $3.15 / 1000Claude Sonnet 5 (Anthropic) — 98.86% · $4.06 / 1000Mistral Small 4 (Mistral) — 98.76% · $0.09 / 1000Gemini 3.5 Flash Lite (Google) — 98.10% · $0.45 / 1000Qwen3.8 Flash (Qwen) — 96.95% · $0.41 / 1000Mistral Small 4Gemma 4 31B100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 99.81% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.81% · $1.40 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.71% · $0.31 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.22 / 1000Claude Sonnet 5.5 (Anthropic) — 99.71% · $3.52 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Gemini 3.8 Flash (Google) — 99.52% · $2.19 / 1000Claude Haiku 5.5 (Anthropic) — 99.43% · $0.21 / 1000GLM-5.3 Flash (Z.ai) — 99.33% · $0.49 / 1000Mistral Large 4 (Mistral) — 99.33% · $3.15 / 1000Claude Sonnet 5 (Anthropic) — 98.86% · $4.06 / 1000Mistral Small 4 (Mistral) — 98.76% · $0.09 / 1000Gemini 3.5 Flash Lite (Google) — 98.10% · $0.45 / 1000Qwen3.8 Flash (Qwen) — 96.95% · $0.41 / 1000Mistral Small 4Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
99.81%$0.12132.9 sec
2Mistral Medium 3.5
MistralEU data
99.81%$1.40130.8 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
99.71%$0.31204.2 sec
4GPT-6 Sol
OpenAIEU data
99.71%$2.22202.5 sec
5Claude Sonnet 5.5New
AnthropicEU data
99.71%$3.52202.2 sec
6GPT-6 Luna
OpenAIEU data
99.62%$0.12271.8 sec
7Gemini 3.8 Flash
Google
99.52%$2.19332.0 sec
8Claude Haiku 5.5New
AnthropicEU data
99.43%$0.21401.8 sec
9GLM-5.3 FlashNew
Z.ai · open weights
99.33%$0.4973.8 sec
10Mistral Large 4New
MistralEU data
99.33%$3.154718.5 sec
11Claude Sonnet 5
AnthropicEU data
98.86%$4.06803.2 sec
12Mistral Small 4
Mistral · open weightsEU data
98.76%$0.09800.9 sec
13Gemini 3.5 Flash Lite
GoogleEU data
98.10%$0.451200.8 sec
14Qwen3.8 Flash
Qwen
96.95%$0.411139.6 sec

Personal data detection · Czech · Text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 99.81% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.81% · $1.40 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.71% · $0.31 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.22 / 1000Claude Sonnet 5.5 (Anthropic) — 99.71% · $3.52 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Gemini 3.8 Flash (Google) — 99.52% · $2.19 / 1000Claude Haiku 5.5 (Anthropic) — 99.43% · $0.21 / 1000GLM-5.3 Flash (Z.ai) — 99.33% · $0.49 / 1000Mistral Large 4 (Mistral) — 99.33% · $3.15 / 1000Claude Sonnet 5 (Anthropic) — 98.86% · $4.06 / 1000Mistral Small 4 (Mistral) — 98.76% · $0.09 / 1000Gemini 3.5 Flash Lite (Google) — 98.10% · $0.45 / 1000Qwen3.8 Flash (Qwen) — 96.95% · $0.41 / 1000Mistral Small 4Gemma 4 31B100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 99.81% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.81% · $1.40 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.71% · $0.31 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.22 / 1000Claude Sonnet 5.5 (Anthropic) — 99.71% · $3.52 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Gemini 3.8 Flash (Google) — 99.52% · $2.19 / 1000Claude Haiku 5.5 (Anthropic) — 99.43% · $0.21 / 1000GLM-5.3 Flash (Z.ai) — 99.33% · $0.49 / 1000Mistral Large 4 (Mistral) — 99.33% · $3.15 / 1000Claude Sonnet 5 (Anthropic) — 98.86% · $4.06 / 1000Mistral Small 4 (Mistral) — 98.76% · $0.09 / 1000Gemini 3.5 Flash Lite (Google) — 98.10% · $0.45 / 1000Qwen3.8 Flash (Qwen) — 96.95% · $0.41 / 1000Mistral Small 4Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
99.81%$0.12132.9 sec
2Mistral Medium 3.5
MistralEU data
99.81%$1.40130.8 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
99.71%$0.31204.2 sec
4GPT-6 Sol
OpenAIEU data
99.71%$2.22202.5 sec
5Claude Sonnet 5.5New
AnthropicEU data
99.71%$3.52202.2 sec
6GPT-6 Luna
OpenAIEU data
99.62%$0.12271.8 sec
7Gemini 3.8 Flash
Google
99.52%$2.19332.0 sec
8Claude Haiku 5.5New
AnthropicEU data
99.43%$0.21401.8 sec
9GLM-5.3 FlashNew
Z.ai · open weights
99.33%$0.4973.8 sec
10Mistral Large 4New
MistralEU data
99.33%$3.154718.5 sec
11Claude Sonnet 5
AnthropicEU data
98.86%$4.06803.2 sec
12Mistral Small 4
Mistral · open weightsEU data
98.76%$0.09800.9 sec
13Gemini 3.5 Flash Lite
GoogleEU data
98.10%$0.451200.8 sec
14Qwen3.8 Flash
Qwen
96.95%$0.411139.6 sec

Personal data detection · German · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Medium 3.5 (Mistral) — 100.00% · $1.38 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.87 / 1000Gemma 4 31B (Google) — 99.90% · $0.10 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.81% · $0.31 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.40 / 1000Gemini 3.8 Flash (Google) — 99.71% · $2.12 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.13 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $4.40 / 1000Qwen3.8 Flash (Qwen) — 99.62% · $0.36 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Claude Haiku 5.5 (Anthropic) — 99.62% · $0.23 / 1000Gemini 3.5 Flash Lite (Google) — 99.62% · $0.45 / 1000Mistral Large 4 (Mistral) — 99.52% · $2.61 / 1000Mistral Small 4 (Mistral) — 99.14% · $0.09 / 1000Mistral Small 4Gemma 4 31BMistral Medium 3.5100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Medium 3.5 (Mistral) — 100.00% · $1.38 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.87 / 1000Gemma 4 31B (Google) — 99.90% · $0.10 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.81% · $0.31 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.40 / 1000Gemini 3.8 Flash (Google) — 99.71% · $2.12 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.13 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $4.40 / 1000Qwen3.8 Flash (Qwen) — 99.62% · $0.36 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Claude Haiku 5.5 (Anthropic) — 99.62% · $0.23 / 1000Gemini 3.5 Flash Lite (Google) — 99.62% · $0.45 / 1000Mistral Large 4 (Mistral) — 99.52% · $2.61 / 1000Mistral Small 4 (Mistral) — 99.14% · $0.09 / 1000Mistral Small 4Gemma 4 31BMistral Medium 3.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Medium 3.5
MistralEU data
100.00%$1.3800.9 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.8701.9 sec
3Gemma 4 31B
Google · open weights
99.90%$0.1071.9 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
99.81%$0.31133.7 sec
5GLM-5.3 FlashNew
Z.ai · open weights
99.81%$0.40133.5 sec
6Gemini 3.8 Flash
Google
99.71%$2.12202.3 sec
7GPT-6 Sol
OpenAIEU data
99.71%$2.13202.6 sec
8Claude Sonnet 5
AnthropicEU data
99.71%$4.40203.7 sec
9Qwen3.8 Flash
Qwen
99.62%$0.36712.2 sec
10GPT-6 Luna
OpenAIEU data
99.62%$0.12272.2 sec
11Claude Haiku 5.5New
AnthropicEU data
99.62%$0.23271.8 sec
12Gemini 3.5 Flash Lite
GoogleEU data
99.62%$0.45270.8 sec
13Mistral Large 4New
MistralEU data
99.52%$2.613318.6 sec
14Mistral Small 4
Mistral · open weightsEU data
99.14%$0.09600.9 sec

Personal data detection · German · Text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Medium 3.5 (Mistral) — 100.00% · $1.38 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.87 / 1000Gemma 4 31B (Google) — 99.90% · $0.10 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.81% · $0.31 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.40 / 1000Gemini 3.8 Flash (Google) — 99.71% · $2.12 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.13 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $4.40 / 1000Qwen3.8 Flash (Qwen) — 99.62% · $0.36 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Claude Haiku 5.5 (Anthropic) — 99.62% · $0.23 / 1000Gemini 3.5 Flash Lite (Google) — 99.62% · $0.45 / 1000Mistral Large 4 (Mistral) — 99.52% · $2.61 / 1000Mistral Small 4 (Mistral) — 99.14% · $0.09 / 1000Mistral Small 4Gemma 4 31BMistral Medium 3.5100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Medium 3.5 (Mistral) — 100.00% · $1.38 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.87 / 1000Gemma 4 31B (Google) — 99.90% · $0.10 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.81% · $0.31 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.40 / 1000Gemini 3.8 Flash (Google) — 99.71% · $2.12 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.13 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $4.40 / 1000Qwen3.8 Flash (Qwen) — 99.62% · $0.36 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Claude Haiku 5.5 (Anthropic) — 99.62% · $0.23 / 1000Gemini 3.5 Flash Lite (Google) — 99.62% · $0.45 / 1000Mistral Large 4 (Mistral) — 99.52% · $2.61 / 1000Mistral Small 4 (Mistral) — 99.14% · $0.09 / 1000Mistral Small 4Gemma 4 31BMistral Medium 3.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Medium 3.5
MistralEU data
100.00%$1.3800.9 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.8701.9 sec
3Gemma 4 31B
Google · open weights
99.90%$0.1071.9 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
99.81%$0.31133.7 sec
5GLM-5.3 FlashNew
Z.ai · open weights
99.81%$0.40133.5 sec
6Gemini 3.8 Flash
Google
99.71%$2.12202.3 sec
7GPT-6 Sol
OpenAIEU data
99.71%$2.13202.6 sec
8Claude Sonnet 5
AnthropicEU data
99.71%$4.40203.7 sec
9Qwen3.8 Flash
Qwen
99.62%$0.36712.2 sec
10GPT-6 Luna
OpenAIEU data
99.62%$0.12272.2 sec
11Claude Haiku 5.5New
AnthropicEU data
99.62%$0.23271.8 sec
12Gemini 3.5 Flash Lite
GoogleEU data
99.62%$0.45270.8 sec
13Mistral Large 4New
MistralEU data
99.52%$2.613318.6 sec
14Mistral Small 4
Mistral · open weightsEU data
99.14%$0.09600.9 sec

Personal data detection · English (US) · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 100.00% · $0.33 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.22 / 1000Gemma 4 31B (Google) — 99.90% · $0.08 / 1000Claude Haiku 5.5 (Anthropic) — 99.81% · $0.19 / 1000Mistral Medium 3.5 (Mistral) — 99.71% · $1.15 / 1000Gemini 3.8 Flash (Google) — 99.71% · $1.86 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.09 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $3.50 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.11 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.62% · $0.29 / 1000Mistral Large 4 (Mistral) — 99.62% · $2.89 / 1000Gemini 3.5 Flash Lite (Google) — 99.52% · $0.41 / 1000Mistral Small 4 (Mistral) — 99.43% · $0.08 / 1000Qwen3.8 Flash (Qwen) — 99.24% · $0.28 / 1000Mistral Small 4Gemma 4 31BGLM-5.3 Flash100%95%90%$0.10$0.50$1Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 100.00% · $0.33 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.22 / 1000Gemma 4 31B (Google) — 99.90% · $0.08 / 1000Claude Haiku 5.5 (Anthropic) — 99.81% · $0.19 / 1000Mistral Medium 3.5 (Mistral) — 99.71% · $1.15 / 1000Gemini 3.8 Flash (Google) — 99.71% · $1.86 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.09 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $3.50 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.11 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.62% · $0.29 / 1000Mistral Large 4 (Mistral) — 99.62% · $2.89 / 1000Gemini 3.5 Flash Lite (Google) — 99.52% · $0.41 / 1000Mistral Small 4 (Mistral) — 99.43% · $0.08 / 1000Qwen3.8 Flash (Qwen) — 99.24% · $0.28 / 1000Gemma 4 31BGLM-5.3 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.3302.8 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.2201.9 sec
3Gemma 4 31B
Google · open weights
99.90%$0.0872.4 sec
4Claude Haiku 5.5New
AnthropicEU data
99.81%$0.19131.7 sec
5Mistral Medium 3.5
MistralEU data
99.71%$1.15200.8 sec
6Gemini 3.8 Flash
Google
99.71%$1.86201.9 sec
7GPT-6 Sol
OpenAIEU data
99.71%$2.09202.6 sec
8Claude Sonnet 5
AnthropicEU data
99.71%$3.50202.5 sec
9GPT-6 Luna
OpenAIEU data
99.62%$0.11271.8 sec
10DeepSeek V4.1 Flash
DeepSeek · open weights
99.62%$0.29273.8 sec
11Mistral Large 4New
MistralEU data
99.62%$2.892728.2 sec
12Gemini 3.5 Flash Lite
GoogleEU data
99.52%$0.41270.8 sec
13Mistral Small 4
Mistral · open weightsEU data
99.43%$0.08400.8 sec
14Qwen3.8 Flash
Qwen
99.24%$0.28137.5 sec

Personal data detection · English (US) · Text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 100.00% · $0.33 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.22 / 1000Gemma 4 31B (Google) — 99.90% · $0.08 / 1000Claude Haiku 5.5 (Anthropic) — 99.81% · $0.19 / 1000Mistral Medium 3.5 (Mistral) — 99.71% · $1.15 / 1000Gemini 3.8 Flash (Google) — 99.71% · $1.86 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.09 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $3.50 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.11 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.62% · $0.29 / 1000Mistral Large 4 (Mistral) — 99.62% · $2.89 / 1000Gemini 3.5 Flash Lite (Google) — 99.52% · $0.41 / 1000Mistral Small 4 (Mistral) — 99.43% · $0.08 / 1000Qwen3.8 Flash (Qwen) — 99.24% · $0.28 / 1000Mistral Small 4Gemma 4 31BGLM-5.3 Flash100%95%90%$0.10$0.50$1Cost per 1,000 documents (log scale)Correct fieldsbar 95%GLM-5.3 Flash (Z.ai) — 100.00% · $0.33 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.22 / 1000Gemma 4 31B (Google) — 99.90% · $0.08 / 1000Claude Haiku 5.5 (Anthropic) — 99.81% · $0.19 / 1000Mistral Medium 3.5 (Mistral) — 99.71% · $1.15 / 1000Gemini 3.8 Flash (Google) — 99.71% · $1.86 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.09 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $3.50 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.11 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.62% · $0.29 / 1000Mistral Large 4 (Mistral) — 99.62% · $2.89 / 1000Gemini 3.5 Flash Lite (Google) — 99.52% · $0.41 / 1000Mistral Small 4 (Mistral) — 99.43% · $0.08 / 1000Qwen3.8 Flash (Qwen) — 99.24% · $0.28 / 1000Gemma 4 31BGLM-5.3 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.3302.8 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.2201.9 sec
3Gemma 4 31B
Google · open weights
99.90%$0.0872.4 sec
4Claude Haiku 5.5New
AnthropicEU data
99.81%$0.19131.7 sec
5Mistral Medium 3.5
MistralEU data
99.71%$1.15200.8 sec
6Gemini 3.8 Flash
Google
99.71%$1.86201.9 sec
7GPT-6 Sol
OpenAIEU data
99.71%$2.09202.6 sec
8Claude Sonnet 5
AnthropicEU data
99.71%$3.50202.5 sec
9GPT-6 Luna
OpenAIEU data
99.62%$0.11271.8 sec
10DeepSeek V4.1 Flash
DeepSeek · open weights
99.62%$0.29273.8 sec
11Mistral Large 4New
MistralEU data
99.62%$2.892728.2 sec
12Gemini 3.5 Flash Lite
GoogleEU data
99.52%$0.41270.8 sec
13Mistral Small 4
Mistral · open weightsEU data
99.43%$0.08400.8 sec
14Qwen3.8 Flash
Qwen
99.24%$0.28137.5 sec

Personal data detection · Spanish · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.23 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.78 / 1000Gemma 4 31B (Google) — 99.90% · $0.10 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.90% · $0.44 / 1000Qwen3.8 Flash (Qwen) — 99.81% · $0.38 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.46 / 1000Gemini 3.8 Flash (Google) — 99.81% · $2.11 / 1000GPT-6 Sol (OpenAI) — 99.81% · $2.21 / 1000Mistral Large 4 (Mistral) — 99.81% · $3.16 / 1000Claude Sonnet 5 (Anthropic) — 99.81% · $4.38 / 1000GPT-6 Luna (OpenAI) — 99.71% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.62% · $1.41 / 1000Mistral Small 4 (Mistral) — 99.05% · $0.09 / 1000Gemini 3.5 Flash Lite (Google) — 95.81% · $0.45 / 1000Mistral Small 4Gemma 4 31BClaude Haiku 5.5100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.23 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.78 / 1000Gemma 4 31B (Google) — 99.90% · $0.10 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.90% · $0.44 / 1000Qwen3.8 Flash (Qwen) — 99.81% · $0.38 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.46 / 1000Gemini 3.8 Flash (Google) — 99.81% · $2.11 / 1000GPT-6 Sol (OpenAI) — 99.81% · $2.21 / 1000Mistral Large 4 (Mistral) — 99.81% · $3.16 / 1000Claude Sonnet 5 (Anthropic) — 99.81% · $4.38 / 1000GPT-6 Luna (OpenAI) — 99.71% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.62% · $1.41 / 1000Mistral Small 4 (Mistral) — 99.05% · $0.09 / 1000Gemini 3.5 Flash Lite (Google) — 95.81% · $0.45 / 1000Mistral Small 4Claude Haiku 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Haiku 5.5New
AnthropicEU data
100.00%$0.2301.9 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.7801.9 sec
3Gemma 4 31B
Google · open weights
99.90%$0.1073.6 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
99.90%$0.4475.9 sec
5Qwen3.8 Flash
Qwen
99.81%$0.381310.2 sec
6GLM-5.3 FlashNew
Z.ai · open weights
99.81%$0.46133.4 sec
7Gemini 3.8 Flash
Google
99.81%$2.11132.3 sec
8GPT-6 Sol
OpenAIEU data
99.81%$2.21132.9 sec
9Mistral Large 4New
MistralEU data
99.81%$3.161321.9 sec
10Claude Sonnet 5
AnthropicEU data
99.81%$4.38133.6 sec
11GPT-6 Luna
OpenAIEU data
99.71%$0.12202.3 sec
12Mistral Medium 3.5
MistralEU data
99.62%$1.41271.0 sec
13Mistral Small 4
Mistral · open weightsEU data
99.05%$0.09670.9 sec
14Gemini 3.5 Flash Lite
GoogleEU data
95.81%$0.452670.9 sec

Personal data detection · Spanish · Text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.23 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.78 / 1000Gemma 4 31B (Google) — 99.90% · $0.10 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.90% · $0.44 / 1000Qwen3.8 Flash (Qwen) — 99.81% · $0.38 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.46 / 1000Gemini 3.8 Flash (Google) — 99.81% · $2.11 / 1000GPT-6 Sol (OpenAI) — 99.81% · $2.21 / 1000Mistral Large 4 (Mistral) — 99.81% · $3.16 / 1000Claude Sonnet 5 (Anthropic) — 99.81% · $4.38 / 1000GPT-6 Luna (OpenAI) — 99.71% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.62% · $1.41 / 1000Mistral Small 4 (Mistral) — 99.05% · $0.09 / 1000Gemini 3.5 Flash Lite (Google) — 95.81% · $0.45 / 1000Mistral Small 4Gemma 4 31BClaude Haiku 5.5100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Haiku 5.5 (Anthropic) — 100.00% · $0.23 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.78 / 1000Gemma 4 31B (Google) — 99.90% · $0.10 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.90% · $0.44 / 1000Qwen3.8 Flash (Qwen) — 99.81% · $0.38 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.46 / 1000Gemini 3.8 Flash (Google) — 99.81% · $2.11 / 1000GPT-6 Sol (OpenAI) — 99.81% · $2.21 / 1000Mistral Large 4 (Mistral) — 99.81% · $3.16 / 1000Claude Sonnet 5 (Anthropic) — 99.81% · $4.38 / 1000GPT-6 Luna (OpenAI) — 99.71% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.62% · $1.41 / 1000Mistral Small 4 (Mistral) — 99.05% · $0.09 / 1000Gemini 3.5 Flash Lite (Google) — 95.81% · $0.45 / 1000Mistral Small 4Claude Haiku 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Haiku 5.5New
AnthropicEU data
100.00%$0.2301.9 sec
2Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.7801.9 sec
3Gemma 4 31B
Google · open weights
99.90%$0.1073.6 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
99.90%$0.4475.9 sec
5Qwen3.8 Flash
Qwen
99.81%$0.381310.2 sec
6GLM-5.3 FlashNew
Z.ai · open weights
99.81%$0.46133.4 sec
7Gemini 3.8 Flash
Google
99.81%$2.11132.3 sec
8GPT-6 Sol
OpenAIEU data
99.81%$2.21132.9 sec
9Mistral Large 4New
MistralEU data
99.81%$3.161321.9 sec
10Claude Sonnet 5
AnthropicEU data
99.81%$4.38133.6 sec
11GPT-6 Luna
OpenAIEU data
99.71%$0.12202.3 sec
12Mistral Medium 3.5
MistralEU data
99.62%$1.41271.0 sec
13Mistral Small 4
Mistral · open weightsEU data
99.05%$0.09670.9 sec
14Gemini 3.5 Flash Lite
GoogleEU data
95.81%$0.452670.9 sec

Personal data detection · French · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.81 / 1000Gemma 4 31B (Google) — 99.81% · $0.12 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.41 / 1000Gemini 3.8 Flash (Google) — 99.81% · $2.13 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.71% · $0.37 / 1000Qwen3.8 Flash (Qwen) — 99.71% · $0.40 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.23 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $4.10 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.62% · $1.43 / 1000Claude Haiku 5.5 (Anthropic) — 99.33% · $0.23 / 1000Mistral Small 4 (Mistral) — 98.95% · $0.09 / 1000Mistral Large 4 (Mistral) — 98.67% · $2.96 / 1000Gemini 3.5 Flash Lite (Google) — 94.95% · $0.46 / 1000Mistral Small 4Gemma 4 31BClaude Sonnet 5.5100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.81 / 1000Gemma 4 31B (Google) — 99.81% · $0.12 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.41 / 1000Gemini 3.8 Flash (Google) — 99.81% · $2.13 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.71% · $0.37 / 1000Qwen3.8 Flash (Qwen) — 99.71% · $0.40 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.23 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $4.10 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.62% · $1.43 / 1000Claude Haiku 5.5 (Anthropic) — 99.33% · $0.23 / 1000Mistral Small 4 (Mistral) — 98.95% · $0.09 / 1000Mistral Large 4 (Mistral) — 98.67% · $2.96 / 1000Gemini 3.5 Flash Lite (Google) — 94.95% · $0.46 / 1000Mistral Small 4Gemma 4 31BClaude Sonnet 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.8102.0 sec
2Gemma 4 31B
Google · open weights
99.81%$0.12134.3 sec
3GLM-5.3 FlashNew
Z.ai · open weights
99.81%$0.41133.3 sec
4Gemini 3.8 Flash
Google
99.81%$2.13132.3 sec
5DeepSeek V4.1 Flash
DeepSeek · open weights
99.71%$0.37205.3 sec
6Qwen3.8 Flash
Qwen
99.71%$0.402010.9 sec
7GPT-6 Sol
OpenAIEU data
99.71%$2.23202.9 sec
8Claude Sonnet 5
AnthropicEU data
99.71%$4.10202.5 sec
9GPT-6 Luna
OpenAIEU data
99.62%$0.12272.4 sec
10Mistral Medium 3.5
MistralEU data
99.62%$1.43270.9 sec
11Claude Haiku 5.5New
AnthropicEU data
99.33%$0.23471.8 sec
12Mistral Small 4
Mistral · open weightsEU data
98.95%$0.09671.0 sec
13Mistral Large 4New
MistralEU data
98.67%$2.965323.6 sec
14Gemini 3.5 Flash Lite
GoogleEU data
94.95%$0.463270.8 sec

Personal data detection · French · Text (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.81 / 1000Gemma 4 31B (Google) — 99.81% · $0.12 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.41 / 1000Gemini 3.8 Flash (Google) — 99.81% · $2.13 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.71% · $0.37 / 1000Qwen3.8 Flash (Qwen) — 99.71% · $0.40 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.23 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $4.10 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.62% · $1.43 / 1000Claude Haiku 5.5 (Anthropic) — 99.33% · $0.23 / 1000Mistral Small 4 (Mistral) — 98.95% · $0.09 / 1000Mistral Large 4 (Mistral) — 98.67% · $2.96 / 1000Gemini 3.5 Flash Lite (Google) — 94.95% · $0.46 / 1000Mistral Small 4Gemma 4 31BClaude Sonnet 5.5100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.81 / 1000Gemma 4 31B (Google) — 99.81% · $0.12 / 1000GLM-5.3 Flash (Z.ai) — 99.81% · $0.41 / 1000Gemini 3.8 Flash (Google) — 99.81% · $2.13 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.71% · $0.37 / 1000Qwen3.8 Flash (Qwen) — 99.71% · $0.40 / 1000GPT-6 Sol (OpenAI) — 99.71% · $2.23 / 1000Claude Sonnet 5 (Anthropic) — 99.71% · $4.10 / 1000GPT-6 Luna (OpenAI) — 99.62% · $0.12 / 1000Mistral Medium 3.5 (Mistral) — 99.62% · $1.43 / 1000Claude Haiku 5.5 (Anthropic) — 99.33% · $0.23 / 1000Mistral Small 4 (Mistral) — 98.95% · $0.09 / 1000Mistral Large 4 (Mistral) — 98.67% · $2.96 / 1000Gemini 3.5 Flash Lite (Google) — 94.95% · $0.46 / 1000Mistral Small 4Gemma 4 31BClaude Sonnet 5.5
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.8102.0 sec
2Gemma 4 31B
Google · open weights
99.81%$0.12134.3 sec
3GLM-5.3 FlashNew
Z.ai · open weights
99.81%$0.41133.3 sec
4Gemini 3.8 Flash
Google
99.81%$2.13132.3 sec
5DeepSeek V4.1 Flash
DeepSeek · open weights
99.71%$0.37205.3 sec
6Qwen3.8 Flash
Qwen
99.71%$0.402010.9 sec
7GPT-6 Sol
OpenAIEU data
99.71%$2.23202.9 sec
8Claude Sonnet 5
AnthropicEU data
99.71%$4.10202.5 sec
9GPT-6 Luna
OpenAIEU data
99.62%$0.12272.4 sec
10Mistral Medium 3.5
MistralEU data
99.62%$1.43270.9 sec
11Claude Haiku 5.5New
AnthropicEU data
99.33%$0.23471.8 sec
12Mistral Small 4
Mistral · open weightsEU data
98.95%$0.09671.0 sec
13Mistral Large 4New
MistralEU data
98.67%$2.965323.6 sec
14Gemini 3.5 Flash Lite
GoogleEU data
94.95%$0.463270.8 sec

Document summaries · Czech · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.19 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.21 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.78 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.13 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.31 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.72 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.69 / 1000Gemini 3.5 Flash Lite (Google) — 99.56% · $0.69 / 1000Mistral Small 4 (Mistral) — 99.00% · $0.22 / 1000Claude Sonnet 5.5 (Anthropic) — 98.67% · $5.01 / 1000Claude Haiku 5.5 (Anthropic) — 95.89% · $0.32 / 1000Qwen3.8 Flash (Qwen) — 95.44% · $1.48 / 1000Claude Sonnet 5 (Anthropic) — 93.44% · $6.48 / 1000Mistral Large 4 (Mistral) — 83.33% · $10.72 / 1000Gemma 4 31B100%95%90%85%80%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.19 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.21 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.78 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.13 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.31 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.72 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.69 / 1000Gemini 3.5 Flash Lite (Google) — 99.56% · $0.69 / 1000Mistral Small 4 (Mistral) — 99.00% · $0.22 / 1000Claude Sonnet 5.5 (Anthropic) — 98.67% · $5.01 / 1000Claude Haiku 5.5 (Anthropic) — 95.89% · $0.32 / 1000Qwen3.8 Flash (Qwen) — 95.44% · $1.48 / 1000Claude Sonnet 5 (Anthropic) — 93.44% · $6.48 / 1000Mistral Large 4 (Mistral) — 83.33% · $10.72 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.1903.4 sec
2GPT-6 Luna
OpenAIEU data
100.00%$0.2103.4 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.78013.4 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$1.13010.4 sec
5Mistral Medium 3.5
MistralEU data
100.00%$2.3101.4 sec
6GPT-6 Sol
OpenAIEU data
100.00%$4.7207.6 sec
7Gemini 3.8 Flash
Google
100.00%$6.6906.2 sec
8Gemini 3.5 Flash Lite
GoogleEU data
99.56%$0.69271.0 sec
9Mistral Small 4
Mistral · open weightsEU data
99.00%$0.22601.6 sec
10Claude Sonnet 5.5New
AnthropicEU data
98.67%$5.01802.7 sec
11Claude Haiku 5.5New
AnthropicEU data
95.89%$0.322272.1 sec
12Qwen3.8 Flash
Qwen
95.44%$1.486734.2 sec
13Claude Sonnet 5
AnthropicEU data
93.44%$6.483204.7 sec
14Mistral Large 4New
MistralEU data
83.33%$10.7216755.4 sec

Document summaries · Czech · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%80%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.19 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.21 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.78 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.13 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.31 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.72 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.69 / 1000Gemini 3.5 Flash Lite (Google) — 99.56% · $0.69 / 1000Mistral Small 4 (Mistral) — 99.00% · $0.22 / 1000Claude Sonnet 5.5 (Anthropic) — 98.67% · $5.01 / 1000Claude Haiku 5.5 (Anthropic) — 95.89% · $0.32 / 1000Qwen3.8 Flash (Qwen) — 95.44% · $1.48 / 1000Claude Sonnet 5 (Anthropic) — 93.44% · $6.48 / 1000Mistral Large 4 (Mistral) — 83.33% · $10.72 / 1000Gemma 4 31B100%95%90%85%80%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.19 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.21 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.78 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.13 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.31 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.72 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.69 / 1000Gemini 3.5 Flash Lite (Google) — 99.56% · $0.69 / 1000Mistral Small 4 (Mistral) — 99.00% · $0.22 / 1000Claude Sonnet 5.5 (Anthropic) — 98.67% · $5.01 / 1000Claude Haiku 5.5 (Anthropic) — 95.89% · $0.32 / 1000Qwen3.8 Flash (Qwen) — 95.44% · $1.48 / 1000Claude Sonnet 5 (Anthropic) — 93.44% · $6.48 / 1000Mistral Large 4 (Mistral) — 83.33% · $10.72 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.1903.4 sec
2GPT-6 Luna
OpenAIEU data
100.00%$0.2103.4 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.78013.4 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$1.13010.4 sec
5Mistral Medium 3.5
MistralEU data
100.00%$2.3101.4 sec
6GPT-6 Sol
OpenAIEU data
100.00%$4.7207.6 sec
7Gemini 3.8 Flash
Google
100.00%$6.6906.2 sec
8Gemini 3.5 Flash Lite
GoogleEU data
99.56%$0.69271.0 sec
9Mistral Small 4
Mistral · open weightsEU data
99.00%$0.22601.6 sec
10Claude Sonnet 5.5New
AnthropicEU data
98.67%$5.01802.7 sec
11Claude Haiku 5.5New
AnthropicEU data
95.89%$0.322272.1 sec
12Qwen3.8 Flash
Qwen
95.44%$1.486734.2 sec
13Claude Sonnet 5
AnthropicEU data
93.44%$6.483204.7 sec
14Mistral Large 4New
MistralEU data
83.33%$10.7216755.4 sec

Document summaries · German · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.16 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.62 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.08 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.98 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.11 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.11 / 1000Mistral Small 4 (Mistral) — 99.89% · $0.19 / 1000Gemini 3.5 Flash Lite (Google) — 99.89% · $0.59 / 1000Claude Haiku 5.5 (Anthropic) — 99.44% · $0.46 / 1000Mistral Large 4 (Mistral) — 99.33% · $7.19 / 1000Claude Sonnet 5 (Anthropic) — 99.11% · $5.97 / 1000Claude Sonnet 5.5 (Anthropic) — 99.11% · $5.82 / 1000Qwen3.8 Flash (Qwen) — 95.33% · $1.43 / 1000Gemma 4 31B100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.16 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.62 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.08 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.98 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.11 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.11 / 1000Mistral Small 4 (Mistral) — 99.89% · $0.19 / 1000Gemini 3.5 Flash Lite (Google) — 99.89% · $0.59 / 1000Claude Haiku 5.5 (Anthropic) — 99.44% · $0.46 / 1000Mistral Large 4 (Mistral) — 99.33% · $7.19 / 1000Claude Sonnet 5 (Anthropic) — 99.11% · $5.97 / 1000Claude Sonnet 5.5 (Anthropic) — 99.11% · $5.82 / 1000Qwen3.8 Flash (Qwen) — 95.33% · $1.43 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.1602.9 sec
2GPT-6 Luna
OpenAIEU data
100.00%$0.1903.3 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.6209.8 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$1.0808.4 sec
5Mistral Medium 3.5
MistralEU data
100.00%$1.9801.1 sec
6GPT-6 Sol
OpenAIEU data
100.00%$4.1107.1 sec
7Gemini 3.8 Flash
Google
100.00%$6.1105.8 sec
8Mistral Small 4
Mistral · open weightsEU data
99.89%$0.1971.3 sec
9Gemini 3.5 Flash Lite
GoogleEU data
99.89%$0.5970.9 sec
10Claude Haiku 5.5New
AnthropicEU data
99.44%$0.46333.5 sec
11Mistral Large 4New
MistralEU data
99.33%$7.19748.0 sec
12Claude Sonnet 5
AnthropicEU data
99.11%$5.97403.6 sec
13Claude Sonnet 5.5New
AnthropicEU data
99.11%$5.82532.8 sec
14Qwen3.8 Flash
Qwen
95.33%$1.434729.5 sec

Document summaries · German · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.16 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.62 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.08 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.98 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.11 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.11 / 1000Mistral Small 4 (Mistral) — 99.89% · $0.19 / 1000Gemini 3.5 Flash Lite (Google) — 99.89% · $0.59 / 1000Claude Haiku 5.5 (Anthropic) — 99.44% · $0.46 / 1000Mistral Large 4 (Mistral) — 99.33% · $7.19 / 1000Claude Sonnet 5 (Anthropic) — 99.11% · $5.97 / 1000Claude Sonnet 5.5 (Anthropic) — 99.11% · $5.82 / 1000Qwen3.8 Flash (Qwen) — 95.33% · $1.43 / 1000Gemma 4 31B100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.16 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.62 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.08 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.98 / 1000GPT-6 Sol (OpenAI) — 100.00% · $4.11 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.11 / 1000Mistral Small 4 (Mistral) — 99.89% · $0.19 / 1000Gemini 3.5 Flash Lite (Google) — 99.89% · $0.59 / 1000Claude Haiku 5.5 (Anthropic) — 99.44% · $0.46 / 1000Mistral Large 4 (Mistral) — 99.33% · $7.19 / 1000Claude Sonnet 5 (Anthropic) — 99.11% · $5.97 / 1000Claude Sonnet 5.5 (Anthropic) — 99.11% · $5.82 / 1000Qwen3.8 Flash (Qwen) — 95.33% · $1.43 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.1602.9 sec
2GPT-6 Luna
OpenAIEU data
100.00%$0.1903.3 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.6209.8 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$1.0808.4 sec
5Mistral Medium 3.5
MistralEU data
100.00%$1.9801.1 sec
6GPT-6 Sol
OpenAIEU data
100.00%$4.1107.1 sec
7Gemini 3.8 Flash
Google
100.00%$6.1105.8 sec
8Mistral Small 4
Mistral · open weightsEU data
99.89%$0.1971.3 sec
9Gemini 3.5 Flash Lite
GoogleEU data
99.89%$0.5970.9 sec
10Claude Haiku 5.5New
AnthropicEU data
99.44%$0.46333.5 sec
11Mistral Large 4New
MistralEU data
99.33%$7.19748.0 sec
12Claude Sonnet 5
AnthropicEU data
99.11%$5.97403.6 sec
13Claude Sonnet 5.5New
AnthropicEU data
99.11%$5.82532.8 sec
14Qwen3.8 Flash
Qwen
95.33%$1.434729.5 sec

Document summaries · English (US) · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.13 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.16 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.54 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.82 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.57 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.60 / 1000Gemini 3.8 Flash (Google) — 100.00% · $5.46 / 1000Mistral Large 4 (Mistral) — 100.00% · $5.62 / 1000Qwen3.8 Flash (Qwen) — 99.89% · $0.75 / 1000Mistral Small 4 (Mistral) — 99.67% · $0.14 / 1000GPT-6 Sol (OpenAI) — 99.56% · $3.85 / 1000Gemini 3.5 Flash Lite (Google) — 99.22% · $0.46 / 1000Claude Haiku 5.5 (Anthropic) — 98.33% · $0.34 / 1000Claude Sonnet 5 (Anthropic) — 96.00% · $4.08 / 1000Gemma 4 31B100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.13 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.16 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.54 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.82 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.57 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.60 / 1000Gemini 3.8 Flash (Google) — 100.00% · $5.46 / 1000Mistral Large 4 (Mistral) — 100.00% · $5.62 / 1000Qwen3.8 Flash (Qwen) — 99.89% · $0.75 / 1000Mistral Small 4 (Mistral) — 99.67% · $0.14 / 1000GPT-6 Sol (OpenAI) — 99.56% · $3.85 / 1000Gemini 3.5 Flash Lite (Google) — 99.22% · $0.46 / 1000Claude Haiku 5.5 (Anthropic) — 98.33% · $0.34 / 1000Claude Sonnet 5 (Anthropic) — 96.00% · $4.08 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.1302.2 sec
2GPT-6 Luna
OpenAIEU data
100.00%$0.1602.9 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.5408.8 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.8207.5 sec
5Mistral Medium 3.5
MistralEU data
100.00%$1.5700.9 sec
6Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.6001.8 sec
7Gemini 3.8 Flash
Google
100.00%$5.4605.4 sec
8Mistral Large 4New
MistralEU data
100.00%$5.62037.2 sec
9Qwen3.8 Flash
Qwen
99.89%$0.75717.8 sec
10Mistral Small 4
Mistral · open weightsEU data
99.67%$0.14131.0 sec
11GPT-6 Sol
OpenAIEU data
99.56%$3.85278.1 sec
12Gemini 3.5 Flash Lite
GoogleEU data
99.22%$0.46470.9 sec
13Claude Haiku 5.5New
AnthropicEU data
98.33%$0.341002.7 sec
14Claude Sonnet 5
AnthropicEU data
96.00%$4.082332.9 sec

Document summaries · English (US) · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.10$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.13 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.16 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.54 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.82 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.57 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.60 / 1000Gemini 3.8 Flash (Google) — 100.00% · $5.46 / 1000Mistral Large 4 (Mistral) — 100.00% · $5.62 / 1000Qwen3.8 Flash (Qwen) — 99.89% · $0.75 / 1000Mistral Small 4 (Mistral) — 99.67% · $0.14 / 1000GPT-6 Sol (OpenAI) — 99.56% · $3.85 / 1000Gemini 3.5 Flash Lite (Google) — 99.22% · $0.46 / 1000Claude Haiku 5.5 (Anthropic) — 98.33% · $0.34 / 1000Claude Sonnet 5 (Anthropic) — 96.00% · $4.08 / 1000Gemma 4 31B100%95%90%$0.10$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.13 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.16 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.54 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.82 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.57 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.60 / 1000Gemini 3.8 Flash (Google) — 100.00% · $5.46 / 1000Mistral Large 4 (Mistral) — 100.00% · $5.62 / 1000Qwen3.8 Flash (Qwen) — 99.89% · $0.75 / 1000Mistral Small 4 (Mistral) — 99.67% · $0.14 / 1000GPT-6 Sol (OpenAI) — 99.56% · $3.85 / 1000Gemini 3.5 Flash Lite (Google) — 99.22% · $0.46 / 1000Claude Haiku 5.5 (Anthropic) — 98.33% · $0.34 / 1000Claude Sonnet 5 (Anthropic) — 96.00% · $4.08 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.1302.2 sec
2GPT-6 Luna
OpenAIEU data
100.00%$0.1602.9 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.5408.8 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.8207.5 sec
5Mistral Medium 3.5
MistralEU data
100.00%$1.5700.9 sec
6Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.6001.8 sec
7Gemini 3.8 Flash
Google
100.00%$5.4605.4 sec
8Mistral Large 4New
MistralEU data
100.00%$5.62037.2 sec
9Qwen3.8 Flash
Qwen
99.89%$0.75717.8 sec
10Mistral Small 4
Mistral · open weightsEU data
99.67%$0.14131.0 sec
11GPT-6 Sol
OpenAIEU data
99.56%$3.85278.1 sec
12Gemini 3.5 Flash Lite
GoogleEU data
99.22%$0.46470.9 sec
13Claude Haiku 5.5New
AnthropicEU data
98.33%$0.341002.7 sec
14Claude Sonnet 5
AnthropicEU data
96.00%$4.082332.9 sec

Document summaries · Spanish · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.88 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.86 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.78 / 1000Gemma 4 31B (Google) — 99.67% · $0.16 / 1000GPT-6 Sol (OpenAI) — 97.89% · $5.02 / 1000Mistral Small 4 (Mistral) — 97.78% · $0.18 / 1000GPT-6 Luna (OpenAI) — 97.00% · $0.19 / 1000GLM-5.3 Flash (Z.ai) — 96.67% · $1.54 / 1000Claude Sonnet 5.5 (Anthropic) — 95.56% · $5.55 / 1000Gemini 3.5 Flash Lite (Google) — 95.33% · $0.56 / 1000Qwen3.8 Flash (Qwen) — 94.44% · $1.22 / 1000Claude Sonnet 5 (Anthropic) — 94.22% · $6.32 / 1000Claude Haiku 5.5 (Anthropic) — 94.11% · $0.61 / 1000Mistral Large 4 (Mistral) — 91.33% · $7.99 / 1000Gemma 4 31BDeepSeek V4.1 Flash100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.88 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.86 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.78 / 1000Gemma 4 31B (Google) — 99.67% · $0.16 / 1000GPT-6 Sol (OpenAI) — 97.89% · $5.02 / 1000Mistral Small 4 (Mistral) — 97.78% · $0.18 / 1000GPT-6 Luna (OpenAI) — 97.00% · $0.19 / 1000GLM-5.3 Flash (Z.ai) — 96.67% · $1.54 / 1000Claude Sonnet 5.5 (Anthropic) — 95.56% · $5.55 / 1000Gemini 3.5 Flash Lite (Google) — 95.33% · $0.56 / 1000Qwen3.8 Flash (Qwen) — 94.44% · $1.22 / 1000Claude Sonnet 5 (Anthropic) — 94.22% · $6.32 / 1000Claude Haiku 5.5 (Anthropic) — 94.11% · $0.61 / 1000Mistral Large 4 (Mistral) — 91.33% · $7.99 / 1000Gemma 4 31BDeepSeek V4.1 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.88010.6 sec
2Mistral Medium 3.5
MistralEU data
100.00%$1.8601.0 sec
3Gemini 3.8 Flash
Google
100.00%$6.7805.9 sec
4Gemma 4 31B
Google · open weights
99.67%$0.16202.6 sec
5GPT-6 Sol
OpenAIEU data
97.89%$5.021277.0 sec
6Mistral Small 4
Mistral · open weightsEU data
97.78%$0.181331.2 sec
7GPT-6 Luna
OpenAIEU data
97.00%$0.191803.0 sec
8GLM-5.3 FlashNew
Z.ai · open weights
96.67%$1.54339.2 sec
9Claude Sonnet 5.5New
AnthropicEU data
95.56%$5.552672.1 sec
10Gemini 3.5 Flash Lite
GoogleEU data
95.33%$0.562800.9 sec
11Qwen3.8 Flash
Qwen
94.44%$1.226723.9 sec
12Claude Sonnet 5
AnthropicEU data
94.22%$6.323203.4 sec
13Claude Haiku 5.5New
AnthropicEU data
94.11%$0.613533.8 sec
14Mistral Large 4New
MistralEU data
91.33%$7.998738.3 sec

Document summaries · Spanish · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.88 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.86 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.78 / 1000Gemma 4 31B (Google) — 99.67% · $0.16 / 1000GPT-6 Sol (OpenAI) — 97.89% · $5.02 / 1000Mistral Small 4 (Mistral) — 97.78% · $0.18 / 1000GPT-6 Luna (OpenAI) — 97.00% · $0.19 / 1000GLM-5.3 Flash (Z.ai) — 96.67% · $1.54 / 1000Claude Sonnet 5.5 (Anthropic) — 95.56% · $5.55 / 1000Gemini 3.5 Flash Lite (Google) — 95.33% · $0.56 / 1000Qwen3.8 Flash (Qwen) — 94.44% · $1.22 / 1000Claude Sonnet 5 (Anthropic) — 94.22% · $6.32 / 1000Claude Haiku 5.5 (Anthropic) — 94.11% · $0.61 / 1000Mistral Large 4 (Mistral) — 91.33% · $7.99 / 1000Gemma 4 31BDeepSeek V4.1 Flash100%95%90%85%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.88 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.86 / 1000Gemini 3.8 Flash (Google) — 100.00% · $6.78 / 1000Gemma 4 31B (Google) — 99.67% · $0.16 / 1000GPT-6 Sol (OpenAI) — 97.89% · $5.02 / 1000Mistral Small 4 (Mistral) — 97.78% · $0.18 / 1000GPT-6 Luna (OpenAI) — 97.00% · $0.19 / 1000GLM-5.3 Flash (Z.ai) — 96.67% · $1.54 / 1000Claude Sonnet 5.5 (Anthropic) — 95.56% · $5.55 / 1000Gemini 3.5 Flash Lite (Google) — 95.33% · $0.56 / 1000Qwen3.8 Flash (Qwen) — 94.44% · $1.22 / 1000Claude Sonnet 5 (Anthropic) — 94.22% · $6.32 / 1000Claude Haiku 5.5 (Anthropic) — 94.11% · $0.61 / 1000Mistral Large 4 (Mistral) — 91.33% · $7.99 / 1000Gemma 4 31BDeepSeek V4.1 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.88010.6 sec
2Mistral Medium 3.5
MistralEU data
100.00%$1.8601.0 sec
3Gemini 3.8 Flash
Google
100.00%$6.7805.9 sec
4Gemma 4 31B
Google · open weights
99.67%$0.16202.6 sec
5GPT-6 Sol
OpenAIEU data
97.89%$5.021277.0 sec
6Mistral Small 4
Mistral · open weightsEU data
97.78%$0.181331.2 sec
7GPT-6 Luna
OpenAIEU data
97.00%$0.191803.0 sec
8GLM-5.3 FlashNew
Z.ai · open weights
96.67%$1.54339.2 sec
9Claude Sonnet 5.5New
AnthropicEU data
95.56%$5.552672.1 sec
10Gemini 3.5 Flash Lite
GoogleEU data
95.33%$0.562800.9 sec
11Qwen3.8 Flash
Qwen
94.44%$1.226723.9 sec
12Claude Sonnet 5
AnthropicEU data
94.22%$6.323203.4 sec
13Claude Haiku 5.5New
AnthropicEU data
94.11%$0.613533.8 sec
14Mistral Large 4New
MistralEU data
91.33%$7.998738.3 sec

Document summaries · French · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%90%80%70%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.78 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.25 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.87 / 1000Gemini 3.8 Flash (Google) — 100.00% · $7.06 / 1000Gemma 4 31B (Google) — 99.11% · $0.16 / 1000GPT-6 Luna (OpenAI) — 98.78% · $0.20 / 1000Claude Sonnet 5.5 (Anthropic) — 97.89% · $4.89 / 1000GPT-6 Sol (OpenAI) — 97.22% · $4.58 / 1000Claude Sonnet 5 (Anthropic) — 96.78% · $5.04 / 1000Qwen3.8 Flash (Qwen) — 96.33% · $1.46 / 1000Gemini 3.5 Flash Lite (Google) — 94.67% · $0.65 / 1000Mistral Small 4 (Mistral) — 93.22% · $0.19 / 1000Claude Haiku 5.5 (Anthropic) — 92.78% · $0.66 / 1000Mistral Large 4 (Mistral) — 72.67% · $11.26 / 1000Gemma 4 31BDeepSeek V4.1 Flash100%90%80%70%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.78 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.25 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.87 / 1000Gemini 3.8 Flash (Google) — 100.00% · $7.06 / 1000Gemma 4 31B (Google) — 99.11% · $0.16 / 1000GPT-6 Luna (OpenAI) — 98.78% · $0.20 / 1000Claude Sonnet 5.5 (Anthropic) — 97.89% · $4.89 / 1000GPT-6 Sol (OpenAI) — 97.22% · $4.58 / 1000Claude Sonnet 5 (Anthropic) — 96.78% · $5.04 / 1000Qwen3.8 Flash (Qwen) — 96.33% · $1.46 / 1000Gemini 3.5 Flash Lite (Google) — 94.67% · $0.65 / 1000Mistral Small 4 (Mistral) — 93.22% · $0.19 / 1000Claude Haiku 5.5 (Anthropic) — 92.78% · $0.66 / 1000Mistral Large 4 (Mistral) — 72.67% · $11.26 / 1000Gemma 4 31BDeepSeek V4.1 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.78010.5 sec
2GLM-5.3 FlashNew
Z.ai · open weights
100.00%$1.2508.7 sec
3Mistral Medium 3.5
MistralEU data
100.00%$1.8701.0 sec
4Gemini 3.8 Flash
Google
100.00%$7.0606.7 sec
5Gemma 4 31B
Google · open weights
99.11%$0.16533.2 sec
6GPT-6 Luna
OpenAIEU data
98.78%$0.20733.2 sec
7Claude Sonnet 5.5New
AnthropicEU data
97.89%$4.891272.7 sec
8GPT-6 Sol
OpenAIEU data
97.22%$4.581676.1 sec
9Claude Sonnet 5
AnthropicEU data
96.78%$5.041933.1 sec
10Qwen3.8 Flash
Qwen
96.33%$1.465328.1 sec
11Gemini 3.5 Flash Lite
GoogleEU data
94.67%$0.652870.9 sec
12Mistral Small 4
Mistral · open weightsEU data
93.22%$0.193801.3 sec
13Claude Haiku 5.5New
AnthropicEU data
92.78%$0.664334.0 sec
14Mistral Large 4New
MistralEU data
72.67%$11.2627346.4 sec

Document summaries · French · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%90%80%70%$0.20$0.50$1$2$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.78 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.25 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.87 / 1000Gemini 3.8 Flash (Google) — 100.00% · $7.06 / 1000Gemma 4 31B (Google) — 99.11% · $0.16 / 1000GPT-6 Luna (OpenAI) — 98.78% · $0.20 / 1000Claude Sonnet 5.5 (Anthropic) — 97.89% · $4.89 / 1000GPT-6 Sol (OpenAI) — 97.22% · $4.58 / 1000Claude Sonnet 5 (Anthropic) — 96.78% · $5.04 / 1000Qwen3.8 Flash (Qwen) — 96.33% · $1.46 / 1000Gemini 3.5 Flash Lite (Google) — 94.67% · $0.65 / 1000Mistral Small 4 (Mistral) — 93.22% · $0.19 / 1000Claude Haiku 5.5 (Anthropic) — 92.78% · $0.66 / 1000Mistral Large 4 (Mistral) — 72.67% · $11.26 / 1000Gemma 4 31BDeepSeek V4.1 Flash100%90%80%70%$0.50$1$5$10Cost per 1,000 documents (log scale)Correct fieldsbar 95%DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.78 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $1.25 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.87 / 1000Gemini 3.8 Flash (Google) — 100.00% · $7.06 / 1000Gemma 4 31B (Google) — 99.11% · $0.16 / 1000GPT-6 Luna (OpenAI) — 98.78% · $0.20 / 1000Claude Sonnet 5.5 (Anthropic) — 97.89% · $4.89 / 1000GPT-6 Sol (OpenAI) — 97.22% · $4.58 / 1000Claude Sonnet 5 (Anthropic) — 96.78% · $5.04 / 1000Qwen3.8 Flash (Qwen) — 96.33% · $1.46 / 1000Gemini 3.5 Flash Lite (Google) — 94.67% · $0.65 / 1000Mistral Small 4 (Mistral) — 93.22% · $0.19 / 1000Claude Haiku 5.5 (Anthropic) — 92.78% · $0.66 / 1000Mistral Large 4 (Mistral) — 72.67% · $11.26 / 1000Gemma 4 31BDeepSeek V4.1 Flash
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.78010.5 sec
2GLM-5.3 FlashNew
Z.ai · open weights
100.00%$1.2508.7 sec
3Mistral Medium 3.5
MistralEU data
100.00%$1.8701.0 sec
4Gemini 3.8 Flash
Google
100.00%$7.0606.7 sec
5Gemma 4 31B
Google · open weights
99.11%$0.16533.2 sec
6GPT-6 Luna
OpenAIEU data
98.78%$0.20733.2 sec
7Claude Sonnet 5.5New
AnthropicEU data
97.89%$4.891272.7 sec
8GPT-6 Sol
OpenAIEU data
97.22%$4.581676.1 sec
9Claude Sonnet 5
AnthropicEU data
96.78%$5.041933.1 sec
10Qwen3.8 Flash
Qwen
96.33%$1.465328.1 sec
11Gemini 3.5 Flash Lite
GoogleEU data
94.67%$0.652870.9 sec
12Mistral Small 4
Mistral · open weightsEU data
93.22%$0.193801.3 sec
13Claude Haiku 5.5New
AnthropicEU data
92.78%$0.664334.0 sec
14Mistral Large 4New
MistralEU data
72.67%$11.2627346.4 sec

Questions about a document · Czech · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.21 / 1000Mistral Small 4 (Mistral) — 100.00% · $0.21 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.38 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.46 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.58 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.51 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.95 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.03 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.14 / 1000Claude Haiku 5.5 (Anthropic) — 99.78% · $0.30 / 1000Qwen3.8 Flash (Qwen) — 94.22% · $0.71 / 1000Mistral Large 4 (Mistral) — 89.89% · $7.08 / 1000Gemma 4 31B100%95%90%85%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.21 / 1000Mistral Small 4 (Mistral) — 100.00% · $0.21 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.38 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.46 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.58 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.51 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.95 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.03 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.14 / 1000Claude Haiku 5.5 (Anthropic) — 99.78% · $0.30 / 1000Qwen3.8 Flash (Qwen) — 94.22% · $0.71 / 1000Mistral Large 4 (Mistral) — 89.89% · $7.08 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.2101.9 sec
2Mistral Small 4
Mistral · open weightsEU data
100.00%$0.2100.8 sec
3GPT-6 Luna
OpenAIEU data
100.00%$0.2303.1 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3803.5 sec
5GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.4603.0 sec
6Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.5800.8 sec
7Mistral Medium 3.5
MistralEU data
100.00%$2.5100.9 sec
8GPT-6 Sol
OpenAIEU data
100.00%$3.9502.3 sec
9Gemini 3.8 Flash
Google
100.00%$4.3603.6 sec
10Claude Sonnet 5.5New
AnthropicEU data
100.00%$5.0302.1 sec
11Claude Sonnet 5
AnthropicEU data
100.00%$5.1402.6 sec
12Claude Haiku 5.5New
AnthropicEU data
99.78%$0.30131.9 sec
13Qwen3.8 Flash
Qwen
94.22%$0.7115314.3 sec
14Mistral Large 4New
MistralEU data
89.89%$7.0822033.6 sec

Questions about a document · Czech · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%85%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.21 / 1000Mistral Small 4 (Mistral) — 100.00% · $0.21 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.38 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.46 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.58 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.51 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.95 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.03 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.14 / 1000Claude Haiku 5.5 (Anthropic) — 99.78% · $0.30 / 1000Qwen3.8 Flash (Qwen) — 94.22% · $0.71 / 1000Mistral Large 4 (Mistral) — 89.89% · $7.08 / 1000Gemma 4 31B100%95%90%85%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.21 / 1000Mistral Small 4 (Mistral) — 100.00% · $0.21 / 1000GPT-6 Luna (OpenAI) — 100.00% · $0.23 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.38 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.46 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.58 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.51 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.95 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.03 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.14 / 1000Claude Haiku 5.5 (Anthropic) — 99.78% · $0.30 / 1000Qwen3.8 Flash (Qwen) — 94.22% · $0.71 / 1000Mistral Large 4 (Mistral) — 89.89% · $7.08 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.2101.9 sec
2Mistral Small 4
Mistral · open weightsEU data
100.00%$0.2100.8 sec
3GPT-6 Luna
OpenAIEU data
100.00%$0.2303.1 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3803.5 sec
5GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.4603.0 sec
6Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.5800.8 sec
7Mistral Medium 3.5
MistralEU data
100.00%$2.5100.9 sec
8GPT-6 Sol
OpenAIEU data
100.00%$3.9502.3 sec
9Gemini 3.8 Flash
Google
100.00%$4.3603.6 sec
10Claude Sonnet 5.5New
AnthropicEU data
100.00%$5.0302.1 sec
11Claude Sonnet 5
AnthropicEU data
100.00%$5.1402.6 sec
12Claude Haiku 5.5New
AnthropicEU data
99.78%$0.30131.9 sec
13Qwen3.8 Flash
Qwen
94.22%$0.7115314.3 sec
14Mistral Large 4New
MistralEU data
89.89%$7.0822033.6 sec

Questions about a document · German · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.32 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.33 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.52 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.13 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.23 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.76 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.49 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.49 / 1000Qwen3.8 Flash (Qwen) — 98.89% · $0.46 / 1000GLM-5.3 Flash (Z.ai) — 98.78% · $0.41 / 1000GPT-6 Luna (OpenAI) — 98.00% · $0.20 / 1000Mistral Large 4 (Mistral) — 93.56% · $5.77 / 1000Mistral Small 4100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.32 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.33 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.52 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.13 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.23 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.76 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.49 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.49 / 1000Qwen3.8 Flash (Qwen) — 98.89% · $0.46 / 1000GLM-5.3 Flash (Z.ai) — 98.78% · $0.41 / 1000GPT-6 Luna (OpenAI) — 98.00% · $0.20 / 1000Mistral Large 4 (Mistral) — 93.56% · $5.77 / 1000Mistral Small 4
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Small 4
Mistral · open weightsEU data
100.00%$0.1800.8 sec
2Gemma 4 31B
Google · open weights
100.00%$0.1802.1 sec
3Claude Haiku 5.5New
AnthropicEU data
100.00%$0.3201.9 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3303.0 sec
5Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.5200.7 sec
6Mistral Medium 3.5
MistralEU data
100.00%$2.1300.8 sec
7GPT-6 Sol
OpenAIEU data
100.00%$3.2302.2 sec
8Gemini 3.8 Flash
Google
100.00%$3.7603.2 sec
9Claude Sonnet 5.5New
AnthropicEU data
100.00%$5.4901.9 sec
10Claude Sonnet 5
AnthropicEU data
100.00%$5.4902.3 sec
11Qwen3.8 Flash
Qwen
98.89%$0.46338.6 sec
12GLM-5.3 FlashNew
Z.ai · open weights
98.78%$0.41472.3 sec
13GPT-6 Luna
OpenAIEU data
98.00%$0.20873.2 sec
14Mistral Large 4New
MistralEU data
93.56%$5.7710731.5 sec

Questions about a document · German · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.32 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.33 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.52 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.13 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.23 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.76 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.49 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.49 / 1000Qwen3.8 Flash (Qwen) — 98.89% · $0.46 / 1000GLM-5.3 Flash (Z.ai) — 98.78% · $0.41 / 1000GPT-6 Luna (OpenAI) — 98.00% · $0.20 / 1000Mistral Large 4 (Mistral) — 93.56% · $5.77 / 1000Mistral Small 4100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.32 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.33 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.52 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.13 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.23 / 1000Gemini 3.8 Flash (Google) — 100.00% · $3.76 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $5.49 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $5.49 / 1000Qwen3.8 Flash (Qwen) — 98.89% · $0.46 / 1000GLM-5.3 Flash (Z.ai) — 98.78% · $0.41 / 1000GPT-6 Luna (OpenAI) — 98.00% · $0.20 / 1000Mistral Large 4 (Mistral) — 93.56% · $5.77 / 1000Mistral Small 4
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Small 4
Mistral · open weightsEU data
100.00%$0.1800.8 sec
2Gemma 4 31B
Google · open weights
100.00%$0.1802.1 sec
3Claude Haiku 5.5New
AnthropicEU data
100.00%$0.3201.9 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3303.0 sec
5Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.5200.7 sec
6Mistral Medium 3.5
MistralEU data
100.00%$2.1300.8 sec
7GPT-6 Sol
OpenAIEU data
100.00%$3.2302.2 sec
8Gemini 3.8 Flash
Google
100.00%$3.7603.2 sec
9Claude Sonnet 5.5New
AnthropicEU data
100.00%$5.4901.9 sec
10Claude Sonnet 5
AnthropicEU data
100.00%$5.4902.3 sec
11Qwen3.8 Flash
Qwen
98.89%$0.46338.6 sec
12GLM-5.3 FlashNew
Z.ai · open weights
98.78%$0.41472.3 sec
13GPT-6 Luna
OpenAIEU data
98.00%$0.20873.2 sec
14Mistral Large 4New
MistralEU data
93.56%$5.7710731.5 sec

Questions about a document · English (US) · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.15 / 1000Gemma 4 31B (Google) — 100.00% · $0.15 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.23 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.32 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.46 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.81 / 1000GPT-6 Sol (OpenAI) — 100.00% · $2.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.84 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $3.88 / 1000Gemini 3.8 Flash (Google) — 99.89% · $3.00 / 1000Qwen3.8 Flash (Qwen) — 99.78% · $0.44 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.22% · $0.31 / 1000GPT-6 Luna (OpenAI) — 99.11% · $0.15 / 1000Mistral Large 4 (Mistral) — 98.89% · $4.16 / 1000Mistral Small 4100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.15 / 1000Gemma 4 31B (Google) — 100.00% · $0.15 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.23 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.32 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.46 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.81 / 1000GPT-6 Sol (OpenAI) — 100.00% · $2.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.84 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $3.88 / 1000Gemini 3.8 Flash (Google) — 99.89% · $3.00 / 1000Qwen3.8 Flash (Qwen) — 99.78% · $0.44 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.22% · $0.31 / 1000GPT-6 Luna (OpenAI) — 99.11% · $0.15 / 1000Mistral Large 4 (Mistral) — 98.89% · $4.16 / 1000Mistral Small 4
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Small 4
Mistral · open weightsEU data
100.00%$0.1500.8 sec
2Gemma 4 31B
Google · open weights
100.00%$0.1502.1 sec
3Claude Haiku 5.5New
AnthropicEU data
100.00%$0.2301.8 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.3202.4 sec
5Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.4600.8 sec
6Mistral Medium 3.5
MistralEU data
100.00%$1.8100.8 sec
7GPT-6 Sol
OpenAIEU data
100.00%$2.3602.3 sec
8Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.8401.5 sec
9Claude Sonnet 5
AnthropicEU data
100.00%$3.8802.2 sec
10Gemini 3.8 Flash
Google
99.89%$3.0073.0 sec
11Qwen3.8 Flash
Qwen
99.78%$0.44138.7 sec
12DeepSeek V4.1 Flash
DeepSeek · open weights
99.22%$0.31134.2 sec
13GPT-6 Luna
OpenAIEU data
99.11%$0.15403.1 sec
14Mistral Large 4New
MistralEU data
98.89%$4.162717.7 sec

Questions about a document · English (US) · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.15 / 1000Gemma 4 31B (Google) — 100.00% · $0.15 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.23 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.32 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.46 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.81 / 1000GPT-6 Sol (OpenAI) — 100.00% · $2.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.84 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $3.88 / 1000Gemini 3.8 Flash (Google) — 99.89% · $3.00 / 1000Qwen3.8 Flash (Qwen) — 99.78% · $0.44 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.22% · $0.31 / 1000GPT-6 Luna (OpenAI) — 99.11% · $0.15 / 1000Mistral Large 4 (Mistral) — 98.89% · $4.16 / 1000Mistral Small 4100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.15 / 1000Gemma 4 31B (Google) — 100.00% · $0.15 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.23 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.32 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.46 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $1.81 / 1000GPT-6 Sol (OpenAI) — 100.00% · $2.36 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $3.84 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $3.88 / 1000Gemini 3.8 Flash (Google) — 99.89% · $3.00 / 1000Qwen3.8 Flash (Qwen) — 99.78% · $0.44 / 1000DeepSeek V4.1 Flash (DeepSeek) — 99.22% · $0.31 / 1000GPT-6 Luna (OpenAI) — 99.11% · $0.15 / 1000Mistral Large 4 (Mistral) — 98.89% · $4.16 / 1000Mistral Small 4
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Small 4
Mistral · open weightsEU data
100.00%$0.1500.8 sec
2Gemma 4 31B
Google · open weights
100.00%$0.1502.1 sec
3Claude Haiku 5.5New
AnthropicEU data
100.00%$0.2301.8 sec
4GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.3202.4 sec
5Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.4600.8 sec
6Mistral Medium 3.5
MistralEU data
100.00%$1.8100.8 sec
7GPT-6 Sol
OpenAIEU data
100.00%$2.3602.3 sec
8Claude Sonnet 5.5New
AnthropicEU data
100.00%$3.8401.5 sec
9Claude Sonnet 5
AnthropicEU data
100.00%$3.8802.2 sec
10Gemini 3.8 Flash
Google
99.89%$3.0073.0 sec
11Qwen3.8 Flash
Qwen
99.78%$0.44138.7 sec
12DeepSeek V4.1 Flash
DeepSeek · open weights
99.22%$0.31134.2 sec
13GPT-6 Luna
OpenAIEU data
99.11%$0.15403.1 sec
14Mistral Large 4New
MistralEU data
98.89%$4.162717.7 sec

Questions about a document · Spanish · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.29 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.35 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.36 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.18 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.36 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.05 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.89 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.94 / 1000GPT-6 Luna (OpenAI) — 99.67% · $0.19 / 1000Qwen3.8 Flash (Qwen) — 99.67% · $0.53 / 1000Mistral Large 4 (Mistral) — 94.22% · $5.74 / 1000Gemma 4 31B100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.29 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.35 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.36 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.18 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.36 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.05 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.89 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.94 / 1000GPT-6 Luna (OpenAI) — 99.67% · $0.19 / 1000Qwen3.8 Flash (Qwen) — 99.67% · $0.53 / 1000Mistral Large 4 (Mistral) — 94.22% · $5.74 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.1801.9 sec
2Mistral Small 4
Mistral · open weightsEU data
100.00%$0.1800.8 sec
3Claude Haiku 5.5New
AnthropicEU data
100.00%$0.2901.8 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3503.3 sec
5GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.3602.3 sec
6Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.5100.8 sec
7Mistral Medium 3.5
MistralEU data
100.00%$2.1800.8 sec
8GPT-6 Sol
OpenAIEU data
100.00%$3.3602.5 sec
9Gemini 3.8 Flash
Google
100.00%$4.0503.6 sec
10Claude Sonnet 5.5New
AnthropicEU data
100.00%$4.8901.4 sec
11Claude Sonnet 5
AnthropicEU data
100.00%$4.9402.2 sec
12GPT-6 Luna
OpenAIEU data
99.67%$0.19132.9 sec
13Qwen3.8 Flash
Qwen
99.67%$0.531310.3 sec
14Mistral Large 4New
MistralEU data
94.22%$5.749322.8 sec

Questions about a document · Spanish · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.29 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.35 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.36 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.18 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.36 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.05 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.89 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.94 / 1000GPT-6 Luna (OpenAI) — 99.67% · $0.19 / 1000Qwen3.8 Flash (Qwen) — 99.67% · $0.53 / 1000Mistral Large 4 (Mistral) — 94.22% · $5.74 / 1000Gemma 4 31B100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Gemma 4 31B (Google) — 100.00% · $0.18 / 1000Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Claude Haiku 5.5 (Anthropic) — 100.00% · $0.29 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.35 / 1000GLM-5.3 Flash (Z.ai) — 100.00% · $0.36 / 1000Gemini 3.5 Flash Lite (Google) — 100.00% · $0.51 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.18 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.36 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.05 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.89 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.94 / 1000GPT-6 Luna (OpenAI) — 99.67% · $0.19 / 1000Qwen3.8 Flash (Qwen) — 99.67% · $0.53 / 1000Mistral Large 4 (Mistral) — 94.22% · $5.74 / 1000Gemma 4 31B
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Gemma 4 31B
Google · open weights
100.00%$0.1801.9 sec
2Mistral Small 4
Mistral · open weightsEU data
100.00%$0.1800.8 sec
3Claude Haiku 5.5New
AnthropicEU data
100.00%$0.2901.8 sec
4DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3503.3 sec
5GLM-5.3 FlashNew
Z.ai · open weights
100.00%$0.3602.3 sec
6Gemini 3.5 Flash Lite
GoogleEU data
100.00%$0.5100.8 sec
7Mistral Medium 3.5
MistralEU data
100.00%$2.1800.8 sec
8GPT-6 Sol
OpenAIEU data
100.00%$3.3602.5 sec
9Gemini 3.8 Flash
Google
100.00%$4.0503.6 sec
10Claude Sonnet 5.5New
AnthropicEU data
100.00%$4.8901.4 sec
11Claude Sonnet 5
AnthropicEU data
100.00%$4.9402.2 sec
12GPT-6 Luna
OpenAIEU data
99.67%$0.19132.9 sec
13Qwen3.8 Flash
Qwen
99.67%$0.531310.3 sec
14Mistral Large 4New
MistralEU data
94.22%$5.749322.8 sec

Questions about a document · French · All levels (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.36 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.19 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.43 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.06 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.90 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.93 / 1000GLM-5.3 Flash (Z.ai) — 99.67% · $0.39 / 1000GPT-6 Luna (OpenAI) — 99.56% · $0.20 / 1000Gemini 3.5 Flash Lite (Google) — 99.56% · $0.53 / 1000Claude Haiku 5.5 (Anthropic) — 99.22% · $0.28 / 1000Qwen3.8 Flash (Qwen) — 98.78% · $0.57 / 1000Mistral Large 4 (Mistral) — 92.33% · $5.74 / 1000Mistral Small 4100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.36 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.19 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.43 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.06 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.90 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.93 / 1000GLM-5.3 Flash (Z.ai) — 99.67% · $0.39 / 1000GPT-6 Luna (OpenAI) — 99.56% · $0.20 / 1000Gemini 3.5 Flash Lite (Google) — 99.56% · $0.53 / 1000Claude Haiku 5.5 (Anthropic) — 99.22% · $0.28 / 1000Qwen3.8 Flash (Qwen) — 98.78% · $0.57 / 1000Mistral Large 4 (Mistral) — 92.33% · $5.74 / 1000Mistral Small 4
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Small 4
Mistral · open weightsEU data
100.00%$0.1800.8 sec
2Gemma 4 31B
Google · open weights
100.00%$0.1902.1 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3603.6 sec
4Mistral Medium 3.5
MistralEU data
100.00%$2.1900.9 sec
5GPT-6 Sol
OpenAIEU data
100.00%$3.4302.5 sec
6Gemini 3.8 Flash
Google
100.00%$4.0603.7 sec
7Claude Sonnet 5.5New
AnthropicEU data
100.00%$4.9001.3 sec
8Claude Sonnet 5
AnthropicEU data
100.00%$4.9302.1 sec
9GLM-5.3 FlashNew
Z.ai · open weights
99.67%$0.39132.2 sec
10GPT-6 Luna
OpenAIEU data
99.56%$0.20203.0 sec
11Gemini 3.5 Flash Lite
GoogleEU data
99.56%$0.53200.8 sec
12Claude Haiku 5.5New
AnthropicEU data
99.22%$0.28331.8 sec
13Qwen3.8 Flash
Qwen
98.78%$0.572710.4 sec
14Mistral Large 4New
MistralEU data
92.33%$5.7412719.8 sec

Questions about a document · French · Document (150 documents)

Accuracy vs. costTop left is accurate and cheap. The dotted line joins the models no other model beats on both price and accuracy. Click a dot for the model page.
Best accuracy for the price meets below
100%95%90%$0.20$0.50$1$2$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.36 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.19 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.43 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.06 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.90 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.93 / 1000GLM-5.3 Flash (Z.ai) — 99.67% · $0.39 / 1000GPT-6 Luna (OpenAI) — 99.56% · $0.20 / 1000Gemini 3.5 Flash Lite (Google) — 99.56% · $0.53 / 1000Claude Haiku 5.5 (Anthropic) — 99.22% · $0.28 / 1000Qwen3.8 Flash (Qwen) — 98.78% · $0.57 / 1000Mistral Large 4 (Mistral) — 92.33% · $5.74 / 1000Mistral Small 4100%95%90%$0.50$1$5Cost per 1,000 documents (log scale)Correct fieldsbar 95%Mistral Small 4 (Mistral) — 100.00% · $0.18 / 1000Gemma 4 31B (Google) — 100.00% · $0.19 / 1000DeepSeek V4.1 Flash (DeepSeek) — 100.00% · $0.36 / 1000Mistral Medium 3.5 (Mistral) — 100.00% · $2.19 / 1000GPT-6 Sol (OpenAI) — 100.00% · $3.43 / 1000Gemini 3.8 Flash (Google) — 100.00% · $4.06 / 1000Claude Sonnet 5.5 (Anthropic) — 100.00% · $4.90 / 1000Claude Sonnet 5 (Anthropic) — 100.00% · $4.93 / 1000GLM-5.3 Flash (Z.ai) — 99.67% · $0.39 / 1000GPT-6 Luna (OpenAI) — 99.56% · $0.20 / 1000Gemini 3.5 Flash Lite (Google) — 99.56% · $0.53 / 1000Claude Haiku 5.5 (Anthropic) — 99.22% · $0.28 / 1000Qwen3.8 Flash (Qwen) — 98.78% · $0.57 / 1000Mistral Large 4 (Mistral) — 92.33% · $5.74 / 1000Mistral Small 4
#ModelCorrect fieldsThe bar: the share of correct fields below which manual fixes cost more than a cheaper model saves. The black mark on the meter shows it.Per 1,000 documentsTo fix / 1,000How many of 1,000 documents have at least one wrong field and need someone to check and fix them.Speed (median)Details
1Mistral Small 4
Mistral · open weightsEU data
100.00%$0.1800.8 sec
2Gemma 4 31B
Google · open weights
100.00%$0.1902.1 sec
3DeepSeek V4.1 Flash
DeepSeek · open weights
100.00%$0.3603.6 sec
4Mistral Medium 3.5
MistralEU data
100.00%$2.1900.9 sec
5GPT-6 Sol
OpenAIEU data
100.00%$3.4302.5 sec
6Gemini 3.8 Flash
Google
100.00%$4.0603.7 sec
7Claude Sonnet 5.5New
AnthropicEU data
100.00%$4.9001.3 sec
8Claude Sonnet 5
AnthropicEU data
100.00%$4.9302.1 sec
9GLM-5.3 FlashNew
Z.ai · open weights
99.67%$0.39132.2 sec
10GPT-6 Luna
OpenAIEU data
99.56%$0.20203.0 sec
11Gemini 3.5 Flash Lite
GoogleEU data
99.56%$0.53200.8 sec
12Claude Haiku 5.5New
AnthropicEU data
99.22%$0.28331.8 sec
13Qwen3.8 Flash
Qwen
98.78%$0.572710.4 sec
14Mistral Large 4New
MistralEU data
92.33%$5.7412719.8 sec

Log in with email

We use your email only to log you in (and for the digest if you tick it). No spam.