Independent benchmarks of AI models on business work

Which model can do your paperwork, in your language, and what will it really cost?

Every week we run commercial and open models through the same business tasks: invoices, customer emails, personal data and contracts, with summaries and questions about documents next. Every number links to the actual model output and what it really costs, including fixing its mistakes.

Last run September 28, 2026 · 10 models · documents in: English (US), Czech, German

Leaderboards by task

Sorted by accuracy. Costs already include the language tax on tokens. Bar: 95%.

#	Model	Accuracy	Per 1,000	Language tax	Data in the EU
1	GPT-6 Sol OpenAI	99.9%meets the bar	$9.65	×1.00	Yes
2	Claude Sonnet 5 Anthropic	99.8%meets the bar	$8.84	×1.00	Yes
3	Gemini 3.8 Flash Google	99.8%meets the bar	$3.84	×1.00	No
4	GPT-6 Luna OpenAI	99.7%meets the bar	$0.51	×1.00	Yes
5	DeepSeek V4.1 Flash DeepSeek · open weights	99.4%meets the bar	$0.52	×1.00	Self-host only
6	Gemini 3.5 Flash Lite Google	99.2%meets the bar	$0.87	×1.00	Yes
7	Mistral Medium 3.5 Mistral	97.8%meets the bar	$4.37	×1.00	Yes
8	Gemma 4 31B Google · open weights	97.6%meets the bar	$0.17	×1.00	Self-host only
9	Mistral Small 4 Mistral · open weights	95.6%meets the bar	$0.41	—	Yes
10	Qwen3.8 Flash Qwen	82.4%below the bar	$1.15	×1.00	No

#	Model	Accuracy	Per 1,000	Language tax	Data in the EU
1	GPT-6 Sol OpenAI	99.0%meets the bar	$2.44	×1.00	Yes
2	Gemini 3.8 Flash Google	98.7%meets the bar	$2.39	×1.00	No
3	Claude Sonnet 5 Anthropic	98.2%meets the bar	$4.01	×1.00	Yes
4	DeepSeek V4.1 Flash DeepSeek · open weights	97.6%meets the bar	$0.23	×1.00	Self-host only
5	GPT-6 Luna OpenAI	97.5%meets the bar	$0.13	×1.00	Yes
6	Qwen3.8 Flash Qwen	97.5%meets the bar	$0.30	×1.00	No
7	Gemma 4 31B Google · open weights	96.5%meets the bar	$0.12	×1.00	Self-host only
8	Gemini 3.5 Flash Lite Google	96.3%meets the bar	$0.42	×1.00	Yes
9	Mistral Medium 3.5 Mistral	94.9%below the bar	$1.47	×1.00	Yes
10	Mistral Small 4 Mistral · open weights	89.7%below the bar	$0.09	—	Yes

#	Model	Accuracy	Per 1,000	Language tax	Data in the EU
1	Gemma 4 31B Google · open weights	99.9%meets the bar	$0.08	×1.00	Self-host only
2	Claude Sonnet 5 Anthropic	99.7%meets the bar	$3.50	×1.00	Yes
3	Gemini 3.8 Flash Google	99.7%meets the bar	$1.86	×1.00	No
4	Mistral Medium 3.5 Mistral	99.7%meets the bar	$1.15	×1.00	Yes
5	GPT-6 Sol OpenAI	99.7%meets the bar	$2.09	×1.00	Yes
6	DeepSeek V4.1 Flash DeepSeek · open weights	99.6%meets the bar	$0.29	×1.00	Self-host only
7	GPT-6 Luna OpenAI	99.6%meets the bar	$0.11	×1.00	Yes
8	Gemini 3.5 Flash Lite Google	99.5%meets the bar	$0.41	×1.00	Yes
9	Mistral Small 4 Mistral · open weights	99.4%meets the bar	$0.08	—	Yes
10	Qwen3.8 Flash Qwen	99.2%meets the bar	$0.28	×1.00	No

#	Model	Accuracy	Per 1,000	Language tax	Data in the EU
1	Claude Sonnet 5 Anthropic	100.0%meets the bar	$5.04	×1.00	Yes
2	DeepSeek V4.1 Flash DeepSeek · open weights	100.0%meets the bar	$0.32	×1.00	Self-host only
3	GPT-6 Sol OpenAI	100.0%meets the bar	$3.33	×1.00	Yes
4	Gemini 3.5 Flash Lite Google	100.0%meets the bar	$0.69	×1.00	Yes
5	Gemini 3.8 Flash Google	99.9%meets the bar	$2.47	×1.00	No
6	GPT-6 Luna OpenAI	99.8%meets the bar	$0.20	×1.00	Yes
7	Gemma 4 31B Google · open weights	99.6%meets the bar	$0.18	×1.00	Self-host only
8	Mistral Medium 3.5 Mistral	98.1%meets the bar	$2.35	×1.00	Yes
9	Qwen3.8 Flash Qwen	95.1%meets the bar	$0.62	×1.00	No
10	Mistral Small 4 Mistral · open weights	91.5%below the bar	$0.17	—	Yes

Full leaderboard and filters →

Which model can do your paperwork, in your language, and what will it really cost?

Recommender

Leaderboards by task