
Mistral Large 4: an open 1T-parameter model for code and agents
Mistral Large 4: 1T parameters, 1M context, USD 0.70 / 2.10 per million tokens, the benchmarks Mistral published and how to read them, as of November 15, 2026.
Data is as of November 15, 2026 and comes from Mistral pages. Prices and the license may change.
What is Mistral Large 4 and when did it come out?
Mistral announced the public preview of the API on October 6, 2026 and promised the weights for the end of that month. A second post presents the open-weights version and raises the numbers on several tests compared with the preview.
| Fact | Value |
|---|---|
| API ID | mistral-large-4 |
| Weights | mistralai/Mistral-Large-4-1T-A52B (Hugging Face) |
| Date | Preview on October 6, 2026; the model card in Mistral’s documentation says November 5, 2026 |
| Parameters | ~1T total, 52B active (49B per token excluding embeddings and output layers, per Hugging Face) |
| Context | 1M tokens |
| Price | USD 0.70 input / USD 2.10 output per million tokens |
| Open or closed | Open weights, under Mistral’s license (modified MIT) |
What is it good for?
Mistral positions it for agentic coding (agents that edit a repository and run commands) and cybersecurity. Among its API features, the documentation lists function calling, structured outputs, document Q&A, batching and agents. Because it accepts images, it can also describe screenshots.
For a designer or product designer, the useful part is this: it is a low-price option for agents that repeat work (auditing pages, reviewing copy, generating code variants) and, because the weights are open, it can be hosted on your own infrastructure; besides Mistral’s own API, the open-weights post names OpenRouter, Microsoft Foundry and NVIDIA endpoints. If you want to understand what that means, read open vs. closed models.
Which benchmarks did it publish and how do you read them?
Mistral reports these scores in its post; some come from third-party tests (Cognition, Artificial Analysis) that Mistral cites. None of them supports a ranking against models from other labs. For the method, read how to read AI benchmarks.
| Benchmark | Score | Who measured or published it |
|---|---|---|
| Terminal-Bench 4.0 | 41% (28.3% in the preview) | Reported by Mistral; it does not say who ran it |
| BrowseComp | 84.5% | Reported by Mistral; it does not say who measured it |
| Cognition FrontierCode 2.0 | 19.7% | Cognition’s private benchmark, reported by Mistral |
| Artificial Analysis Coding Agent Index | 48% | Artificial Analysis, reported by Mistral |
| Artificial Analysis Cyber Index | 51% | Artificial Analysis, reported by Mistral |
| B3 Agent Security Benchmark | 97.1% of attacks resisted | Reported by Mistral (93.3% in the preview, where it credits Lakera’s public B3) |
How to read them:
- Mistral describes FrontierCode 2.0 as an evaluation of end-to-end code quality and of whether a repository maintainer would accept the agent’s change.
- It describes the Cyber Index as the ability to find vulnerabilities in a codebase, reproduce them, validate them and fix them without breaking existing behavior.
- For BrowseComp, the post gives only the score and does not explain its method. For B3, it cites resistance to attacks such as system prompt exfiltration or malicious code injection.
- The improvement between preview and final version (for example, Terminal-Bench from 28.3% to 41%) is declared by the lab itself. It is useful to see the trend, not an audit.
How much does it cost and how much context does it have?
As of November 15, 2026, Mistral’s documentation lists, in dollars per million tokens: 0.70 for input, 0.07 for cached input and 2.10 for output, with 1M tokens of context. To turn this into the real cost of a task, read cost per token, tokens and context window. If you host it yourself, cost is no longer per token but infrastructure.
When should you pick it and when not?
Pick it when:
- You want an open, low-price model for code or review agents.
- You need image input and text output.
- You care about hosting it on your own infrastructure, or using the European deployment that Mistral says it operates end to end.
Avoid it when:
- You need to generate images or audio: output is text only.
- You plan to offer the model to third parties as a service: the license requires a commercial agreement with Mistral.
- The license is a strict legal requirement at your company: there are two different descriptions (modified MIT and Apache 2.0), so ask legal to review with the license file in hand.
Example
Illustrative example: a product team maintains 60 pages of its site and wants to find buttons without accessible text and inconsistent error messages. They write a script that sends each page’s HTML to mistral-large-4 with reasoning_effort set to medium and asks for a list of problems with the affected line.
What to try tomorrow: run only 5 pages, check by hand whether the problems exist, and compute the token cost of those 5. If cost and accuracy work for you, scale to the 60.
Common mistakes
- Quoting the 41% on Terminal-Bench 4.0 as if it were comparable with another lab’s without checking the harness.
- Believing open means free of conditions: a model-as-a-service business needs a license.
- Always using
highreasoning. For daily tasks,mediummay be enough. - Mixing the preview numbers with those of the open-weights version.
- Forgetting that output is text only.