Mistral Large 4 is a one-trillion-parameter AI model from the French lab Mistral, and you can call it through the API right now. It is a public preview, the open weights are promised for the end of October, and the price is low next to most frontier models.
Here is what the launch actually includes, which numbers to trust, and what is still missing.
The quick facts first:
- Announced: October 6, 2026, as a public preview in Mistral Studio
- Size: 1 trillion total parameters, 52 billion active per token
- Price: $1.36 per million input tokens and $4.18 per million output tokens
- Open weights: promised “by the end of the month,” license not yet stated
- Languages: more than 160, including every official EU language
That list is Mistral’s own summary. The rest of this post separates what the company has shown from what it has only claimed.
What Mistral Large 4 actually is
Mistral calls it a hybrid model. One system handles plain instructions, step-by-step reasoning and agent-style tasks, so you do not pick between separate “chat” and “reasoning” versions.
It is a mixture-of-experts design. Only about 52 billion of the 1 trillion parameters work on any given token, which keeps responses faster and cheaper than a dense model of the same size. Some outlets quote 49 billion active parameters instead, so treat the 52 billion figure on Mistral’s own page as the one to use.
It also reads images, not just text. Mistral lists charts, documents, engineering drawings and satellite imagery as strengths. The context window has not been published yet.
Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers. It describes the launch as the first milestone funded by its €3 billion Series D round.
Mistral Large 4 benchmarks: strong, but mostly self-reported
Mistral’s launch page leans hard on coding and security tests. Here are the headline results it published.
| Benchmark | Mistral Large 4 | What it tells you |
|---|---|---|
| DeepSWE v1.1 | 61.7% | Real software-engineering tasks; the main coding claim |
| AutomationBench | 59.9% | How well it completes multi-step business tasks |
| Cybench | 93% of 40 challenges | Security capture-the-flag style problems |
| Vulnerability reproduce-and-patch | 82% | Mistral calls this the highest score reported |
| Surge AI blind coding review | 3.74 | Second of five models, behind Claude Opus 5 at 4.22 |
The last row is the most useful one. It is a blind human rating, and Mistral admits the model lands second, not first.
Most other figures come from Mistral itself, and the page does not explain how every third-party test was run. Wait for independent runs, especially once the weights are public, before you treat any of them as settled.
What the Mistral Large 4 price means for your API bill
At $1.36 in and $4.18 out per million tokens, the preview undercuts the current mid-tier prices we have tracked. Prices below for the other two models come from Local AI Zone’s October tracker.
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Mistral Large 4 (preview) | $1.36 | $4.18 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
Cheaper is not the same as better, and preview pricing can change. If cost is your main worry, our look at Claude Haiku 5.5’s price cut shows how fast the low end is moving too.
Open weights, licenses and the cyber access tier
Mistral describes Large 4 as open-weight, but nothing is downloadable today. The company says weights arrive by the end of October along with more detail on the architecture and post-training.
The license matters as much as the download. “Open-weight” can mean anything from Apache 2.0 to a restricted community license, and Mistral has not said which one this is. Compare that with Cloudflare’s Clef models, which shipped under a clear Apache 2.0 license from day one.
There is also a second, restricted version. Cybersecurity leaders, vetted partners and state authorities get a build with reduced moderation and stronger cyber abilities while Mistral red-teams the model. The public preview does not include that version.
Running a trillion-parameter model yourself will not be a laptop job. Mistral mentions private-cloud and on-premise deployment, which points at companies with serious GPU budgets.
Frequently Asked Questions
Is Mistral Large 4 free to use?
No. The preview API is paid, at $1.36 per million input tokens and $4.18 per million output tokens. The weights will be free to download if Mistral ships them as promised, though you would need to host them yourself.
When will Mistral Large 4 weights be released?
Mistral says by the end of October 2026. It has not given an exact date or a license yet, so check its announcement page.
Is Mistral Large 4 better than Claude or GPT?
Not across the board. On Mistral’s own blind coding comparison it ranked second, behind Claude Opus 5. It claims wins on some legal, finance and security tests, but those results are mostly self-reported.
How do I try Mistral Large 4 today?
Use Mistral Studio at console.mistral.ai. The documentation at docs.mistral.ai lists the preview model and its status.
Our take: worth a test, not a switch
Run a few of your real prompts through the preview, especially coding and document work, because the price is good enough to justify an afternoon of testing. Hold off on moving production traffic until the weights, the license and the context window are public, and until someone outside Mistral has checked the benchmark claims. For anyone who cares where their data is processed, the European hosting option is the most interesting part of this launch.
Leave a Reply