Mistral Large 4 ranks 8th of 13 and costs $5.50 per 1,000 docs
Mistral Large 4 gets 88.99 % of documents fully right at $5.50 per 1,000. It is strong on email triage and weak on invoices and questions over a document.
Published October 7, 2026 · numbers from run 20261007T165728Z-d649b2
Mistral Large 4 is the newest model in our run of 2026-10-07. It lands in 8th place out of 13. It gets 88.99 % of documents fully right, averaged over every task and language, and the API charged $5.50 per 1,000 documents. A document counts as correct only if every field is right.
The short version: it is a clear step up from Mistral Medium 3.5 on invoices and email triage. But it costs about twice as much and loses to several cheaper models.
Where it is strong and where it is not
The per-task picture is uneven. Here are the scores, with Mistral Medium 3.5 and the best model in each task for reference.
| Task | Mistral Large 4 | Mistral Medium 3.5 | Best model |
|---|---|---|---|
| Contract clauses | 92.13 % | 92.53 % | 99.73 % (Gemini 3.8 Flash) |
| Email triage | 85.33 % | 53.73 % | 89.06 % (GLM-5.3 Flash) |
| Invoice extraction | 82.13 % | 73.73 % | 99.59 % (Claude Sonnet 5.5) |
| PII redaction | 96.53 % | 98.26 % | 99.60 % (Claude Sonnet 5.5) |
| Questions over a document | 88.53 % | 100.00 % | 100.00 % (several) |
| Meeting summaries | 89.33 % | 100.00 % | 100.00 % (several) |
Email triage is its best result relative to the field. At 85.33 % it ranks 4th of 13, behind GLM-5.3 Flash, Gemini 3.8 Flash and GPT-6 Sol. Mistral Medium 3.5 scores 53.73 % on the same task.
The weak spot is questions over a document. Large 4 gets 88.53 % at $5.70 per 1,000. Mistral Medium 3.5, Mistral Small 4 and Gemma 4 31B all reach 100.00 %, and the last two cost under $0.20. Meeting summaries are similar: 89.33 % at $8.56, the highest price of any task for this model, while Mistral Medium 3.5 reaches 100.00 % at $1.92.
Contract clauses and PII redaction are solid but not leading. Mistral Medium 3.5 is slightly ahead on both, at a lower price.
Invoices and scans
Invoice extraction is where the gap to the top is widest. Large 4 gets 95.08 % of fields right, but only 82.13 % of invoices have all 13 fields right. Its weakest fields are bank_account at 89.46 % and vat_id at 91.83 %.
On scanned and photographed invoices it drops to 76.22 %, which is 9th of 13. Claude Sonnet 5.5 gets 100.00 % and Gemini 3.8 Flash gets 98.44 % on the same set. Mistral Medium 3.5 manages 61.11 %, so Large 4 is a real improvement within the Mistral line.
The cost of errors matters more than the API price here. We assume 3 minutes of fixing per wrong invoice at $40 per hour. With that, Large 4 costs $364.41 per 1,000 invoices, API fee included. Gemini 3.8 Flash comes to $48.51 and GLM-5.3 Flash to $110.32. Mistral Medium 3.5 comes to $529.99. If you process invoices, the cheap API price of a weaker model is not the real price. We explain why in field vs document accuracy.
Languages and cost
By document language, Large 4 is most reliable in German (92.33 %) and Spanish (91.77 %). Czech (89.11 %) and English (89.22 %) are close together. French is the weak one at 82.55 %, well below its other languages.
One point in its favour: it needs the fewest tokens for the same document in German, Spanish and French. The factor against English is 1.23x for German, 1.26x for Spanish and 1.28x for French. That is the low end of the range across models. Polish is the opposite: 2.01x, the highest we measured.
At $5.50 per 1,000 documents, it sits in the same price band as the top commercial models, but with lower accuracy:
- Claude Sonnet 5 scores 91.28 % at $5.96.
- Claude Sonnet 5.5 scores 95.48 % at $5.55.
- GLM-5.3 Flash scores 96.55 % at $0.68, about 8x cheaper.
- GPT-6 Luna scores 92.44 % at $0.24, more than 20x cheaper.
- Qwen3.8 Flash scores 88.60 % at $0.74, close to Large 4 for a fraction of the price.
The closest model in the Mistral line is Mistral Medium 3.5, at 86.37 % and $2.45. Large 4 is ahead overall, but it loses to Medium on four of the six tasks or ties it, namely contracts, PII, questions and summaries. Its lead comes from invoices and email triage.
Why a cheap price list still costs $5.50
Its list price is low: $0.68 per million input tokens and $2.09 per million output tokens, less than half the input price and about a quarter of the output price of Mistral Medium 3.5 ($1.50 / $7.50). Yet per document it costs about twice as much. The reason is reasoning: before it answers, Large 4 writes a long chain of thought, and every token of it is billed as output.
On long documents it sometimes runs out of room before the answer: 10.66 % of meeting summaries, 4.66 % of answers to questions over a document and 3.60 % of invoices came back without a valid answer. Mistral Medium 3.5 and Claude Sonnet 5.5 had 0.00 % on all three. Look at the cost per document, not the price per token.
Who should use it
Large 4 makes sense if you need email triage or invoice extraction from the Mistral family and cannot move to another vendor. For most other jobs, a cheaper model in this run is at least as accurate. For invoices, especially scans, the models above it are clearly more reliable.
See all 13 models in the leaderboard. How we test, score and price documents is described in the methodology.