- Mistral Large 4 is a mixture-of-experts model with 1.05T total and 52B active parameters, a 1M token context and image input.
- The preview API lists $0.70 input, $0.07 cached input and $2.10 output per million tokens, far below the big closed models.
- Its strongest claims are cybersecurity, agentic coding and visual grounding; most numbers are Mistral’s own, so test before you commit.
- It is still a public preview: weights were promised for the end of October, so check the licence before you plan to self-host.
Mistral Large 4 is the French lab’s biggest model so far, announced on October 6, 2026 as a public preview. Mistral calls it “le Chonk”, and the name fits: about one trillion parameters, a million-token context window and open weights promised soon.
The question most teams ask is simple. Can a model that costs a fraction of the closed flagships replace them for real work? The honest answer today is “maybe, for some jobs, and you should measure it yourself.”
This Mistral Large 4 review sticks to what Mistral’s own announcement and docs confirm, separates vendor benchmarks from independent ones, and ends with a short test plan you can run in an afternoon.
Mistral Large 4 specs
| Item | What Mistral says |
|---|---|
| Status | Public preview (October 6, 2026), version 26.10 |
| Architecture | Granular mixture of experts |
| Parameters | 1.05T total, 52B active, plus a 1.6B vision encoder |
| Input | Text and images |
| Context window | 1M tokens |
| Training | From scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centres |
| Languages | More than 160 in the training data |
| Licence | Listed as “Mistral Ecosystem Open Weight License” |
Mixture of experts means only part of the network runs for each token. That is why a one-trillion-parameter model can be served at a price closer to a mid-sized one. Some third-party pages quote 49B active parameters, but the official docs say 52B, so use that figure.
Mistral Large 4 pricing
The docs list $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Some aggregator sites show slightly different decimals or temporary sale prices, so check the docs page on the day you budget.
| Model | Input / output per MTok |
|---|---|
| Mistral Large 4 (preview) | $0.70 / $2.10 |
| Claude Sonnet 5.5 | $2 / $10 |
| Claude Opus 5.5 | $4 / $20 |
The Claude figures come from Anthropic’s docs; see our Claude Sonnet 5.5 migration guide. Price is only half of cost, though. A cheaper model that needs two attempts per task, or longer outputs, can end up costing the same. Track cost per finished task, not cost per token.
Where Mistral Large 4 looks strong
These results are reported by Mistral in its launch post unless noted. Treat them as claims to verify.
- Cybersecurity: Mistral says it ranks in the top five on the Artificial Analysis Cyber Index and scores 93% on Cybench. It also says some closed models score near zero on a vulnerability reproduction test because they refuse the task.
- Agentic coding: 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index score of 49.8%.
- Human preference for code: in a blind rating by Surge AI it placed second of five models, behind Claude Opus 5.
- Vision: strong visual grounding, including a 42% vs 41% result against GPT-6 Astra on Dense 200.
- Safety: 93.3% attack resistance on Lakera’s B3 benchmark, according to Mistral.
Notice what is missing: head-to-head wins over the very best closed coding models. Mistral’s own coding chart puts Claude Opus 5 ahead on human ratings. Independent leaderboards that aggregate many tasks also rank it lower than the launch material suggests, so a general “better than GPT” claim is not supported by what is public today.
Open weights: what is confirmed
Mistral says the weights will be released by the end of October, and until then it is red-teaming the model with security partners. At the time of writing I could not confirm that the weights are published, so do not build a self-hosting plan on them yet. When they do land, read the licence carefully. The docs call it the Mistral Ecosystem Open Weight License, which is not the same as Apache 2.0 or MIT, so commercial limits may apply.
Running a model with 52B active parameters also needs serious hardware, because all 1.05T parameters must be stored even if few are used per token. For most teams the API is the practical route.
Trying the Mistral Large 4 API
The model works through the usual chat completions endpoint. The docs list structured outputs, function calling, document QnA, batching and agents among the supported features. The model name in the docs is mistral-large-4.
import os, requests
r = requests.post(
"https://api.mistral.ai/v1/chat/completions",
headers={"Authorization": f"Bearer {os.environ['MISTRAL_API_KEY']}"},
json={
"model": "mistral-large-4",
"messages": [{"role": "user", "content": "Summarise this stack trace in 3 bullets: ..."}],
},
timeout=60,
)
print(r.json()["choices"][0]["message"]["content"])
Check the model alias in the official model page before shipping, because preview names can change when the final release arrives.
A one-afternoon test plan
- Collect 30 real prompts from your product: support replies, bug triage, code review and a few image questions.
- Run them through Mistral Large 4 and your current model with identical instructions.
- Have two people score the answers blind on a 1 to 5 scale.
- Record tokens, latency and retries, then compute cost per accepted answer.
- Repeat with your hardest 5 prompts and with prompt-injection attempts.
If you are still deciding how to tune a model for your data, read RAG vs fine-tuning vs prompt engineering first. For another new flagship to compare, see Gemini 4 Argon.
QA angle: a cheaper model is a good candidate for bulk jobs like test-case drafts or log summaries. Keep a stronger model as a fallback for hard cases and review everything that touches production.
Who should use it
| You are | Verdict |
|---|---|
| A team in Europe with data-residency needs | Worth a pilot; Mistral offers a European deployment |
| A security team that hits refusals on closed models | Strong candidate, with proper access controls |
| A startup watching API cost | Pilot on bulk workloads |
| Anyone who needs a stable, final release | Wait for general availability |
Official sources: the Mistral Large 4 announcement and the models overview.
Final thoughts
Mistral Large 4 is a credible open-weight contender with unusual strength in security work and very low prices. It is not yet proven as a drop-in replacement for the top closed models. Run the small test above, watch for the weights and licence, and keep your current model as a fallback until the preview becomes a final release.
Frequently asked questions
What is Mistral Large 4?
It is Mistral AI’s largest model, a multimodal mixture-of-experts model with 1.05T total and 52B active parameters, released as a public preview on October 6, 2026.
How much does Mistral Large 4 cost?
The docs list $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens during the preview.
Are the Mistral Large 4 weights available?
Mistral said weights would be released by the end of October 2026. Check Mistral’s site for the current status and read the licence before commercial use.
What is the context window?
The docs list a 1M token context window, and the model accepts text and image input.
Is it better than GPT or Claude?
Mistral reports leading open-weight results and strong cybersecurity scores, but its own human coding evaluation ranked Claude Opus 5 higher. Test on your own tasks.
Keep learning
Need help with testing, WordPress or SEO?
We help teams ship better software and grow their sites. Tell us what you are working on and we will suggest a clear next step.