When to use Mistral Large 4
Mistral positions Large 4 as a general frontier model among open weights, with particular strength in vision, finance, law, and business workflows. On coding judged by people, it trails Claude Opus 5.What the preview delivers
Vision on documents, charts, and drawings
The model reads dense scenes, mechanical drawings, PDFs, and large geospatial images. On the Dense 200 visual grounding test, Mistral measured 42%, ahead of GPT-6 Astra at 41%.Finance, law, and deliverables
In third-party evaluations cited by Mistral (vals.ai), Large 4 beat GPT-6 Astra on representative finance and legal tasks. On Harvey’s Legal Agent benchmark, it led open-weight models. On AA-Briefcase, which measures long-horizon knowledge work, it reached 1,393 Elo, ahead of DeepSeek V4 Pro.Tool-using workflows
On AutomationBench, 657 business workflows across apps such as Gmail, Sheets, Slack, and Salesforce, the model scored 59.9%, ahead of Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro.Repository coding, with a clear ceiling
Mistral published 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.0. The combined Coding Agent Index was 49.8%, above DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind Surge AI evaluation, annotators scored outputs from 1 to 5 without knowing which model answered. Large 4 placed second of five, at 3.74, ahead of Kimi K3 (3.59), GLM-5.3 (3.60), and GLM-5.2 (3.40), and behind Claude Opus 5 (4.22).What the independent benchmarks show
On Artificial Analysis, the preview scored 38 on the Intelligence Index, ranking 64th among 226 models in the same class. The median for that price range is 26. In that measurement the model generated text at 116 tokens per second, with 1.46 seconds to the first token. Average cost was 1.36 / $4.18), not the promotional price. The same evaluation flags two caveats. The model was verbose: 200 million output tokens on the index, against a median of 81 million. And the window measured there was about 524K tokens, while Mistral’s official card states 1M.The coding, vision, finance, and law figures above were published by Mistral. The Intelligence Index, speed, and cost per task are from Artificial Analysis on the preview. Results depend on the harness, the tools, and the version measured. Mistral said the preview’s reinforcement-learning run was still in progress. Validate in your own workflow before changing a production setup.
Where it doesn’t lead
- On coding judged by people, Claude Opus 5 was clearly ahead (4.22 versus 3.74).
- On Terminal-Bench 4.0, 28.3% is not the level of closed models built for long agentic coding.
- On Artificial Analysis, the preview writes much more than the median for its range. Ask for short answers when the format matters.
- The weights are not published yet. Preview behavior can change before the release.
Where it adds the most value
A drawing or a PDF that has to be checked
The part, the dimension, or the chart lives in the image, not in a text field. Large 4 zooms in, checks, and returns what it saw, with the place in the image it took the answer from. Expected result: a check in which every number or part points back to the visual source.A financial analysis spread across several files
Statements, notes, and reports do not share a format. The model walks the sources, does the math, and separates what is still missing. Expected result: a spreadsheet and a summary in which every value has a source, and missing data is explicit.A workflow that crosses email, a spreadsheet, and a message
The task does not fit in one reply: something has to be read, recorded, and returned as a deliverable. Large 4 was measured on this kind of business workflow. Expected result: the step is recorded in the right tool, and anything that could not be finished comes back as a pending item, not as a success.How to apply it in Tess
1
Pick work that includes a file or an image
Prefer PDFs, spreadsheets, charts, drawings, and flows that cross more than one source. For long, open-ended coding, compare with Opus 5.5.
2
Lock the format and the length
The preview tends to write more than the median. State the line limit, the spreadsheet columns, and what should not appear.
3
Set tools and approvals
Enable only the tools you need. Require confirmation before sending, publishing, or changing data.
4
Ask for the source of every number
Tell the agent to cite the page, the cell, or the region of the image. If the information is not in the material, it should say so.
5
Treat the preview as a preview
Run a small batch before you switch a production flow. The model was still being tuned when Mistral opened the API.
Three playbooks to get started
1. Check a drawing or a chart
Input: the image or PDF and the list of what must be verified.Tools: file reading.
2. A financial spreadsheet with sources
Input: reports, notes, and the business question.Tools: file reading and spreadsheet generation.
3. A flow between email and a spreadsheet
Input: the incoming message, the tracking spreadsheet, and the rule for what to record.Tools: email and spreadsheet, with confirmation before writing.
Limitations and best practices
- The API is in public preview. The weights were announced for the end of October 2026 and are not yet the production artifact.
- The promotional price (2.09) is half the list published on the same card. Confirm the current price before you estimate cost.
- The official card states a 1M-token context. Artificial Analysis measured about 524K on the preview it evaluated. Do not design a flow that depends on the higher ceiling without testing it.
- A long context does not replace selecting information. Send only relevant content.
- On cybersecurity, Mistral reports strong performance and, at the same time, a higher refusal rate on malicious requests than other open models. Defensive work still follows your organization’s policy.
- The model was trained on data in more than 160 languages, including every official language of the European Union. That does not replace reviewing the text in the language you will deliver.
- Review numbers, citations, and files before publishing or sending.
- Keep human approval for financial, legal, clinical, or high-impact decisions.