Mistral Large 4 price on Mistral’s own API is $0.68 per million input tokens and $2.09 per million output tokens right now. That is a sale price for the Paris lab’s new trillion-parameter model. The list rate on the same docs page is twice as high: $1.36 in and $4.18 out.
Mistral announced the public preview on October 6 and nicknamed the model Le Chonk. You can call it today. You still can’t download the weights. That gap, between a discounted preview API and an open-weight drop that hasn’t happened yet, is the story people are searching for.
Mistral Large 4 price: sale vs list on the official page
Mistral’s model page for Mistral Large 4, which we checked on October 12, 2026, lists both columns side by side. Cached input is $0.07 per million tokens on sale and $0.14 at list. The page does not say when the sale ends.
| Per 1M tokens (API) | Sale price | List / “original” price |
|---|---|---|
| Input | $0.68 | $1.36 |
| Cached input | $0.07 | $0.14 |
| Output | $2.09 | $4.18 |
Context window on that page is 1 million tokens. The model ID in the docs is mistral-large-4, tagged public preview, version v26.10. Batch and conversation endpoints are listed as supported; the page does not publish a tokens-per-second speed figure.
1.05 trillion parameters, open weights later
Mistral’s launch post calls Large 4 a “1 trillion-parameter natively multimodal model with 52 billion active parameters.” The docs page is more precise on the totals: 1.05 trillion parameters, 52 billion active, plus a 1.6 billion-parameter vision encoder. Mistral’s Hugging Face placeholder for the release, Mistral-Large-4.0-1T05-A52B, splits the active count further: 49 billion active parameters per token, or 52 billion including embeddings and output layers. That is company documentation, not an independent measurement.
The weights are not downloadable yet. Mistral’s blog says they “drop end of this month.” The Hugging Face page, checked October 12, shows a current ETA of October 31, 2026. The license for those weights has not been published. Until they ship, the only host is Mistral’s own European infrastructure, which the company says trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs.
What Mistral claims on benchmarks, and what it doesn’t
Mistral’s own tables put Large 4 ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max on a combined Coding Agent Index (49.8%), with 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. On Artificial Analysis AutomationBench it reports 59.9%. On visual grounding it says Large 4 scored 42% on Dense 200 versus 41% for GPT-6 Astra. Those are Mistral’s runs and Mistral’s framing. The blog also says Large 4 “significantly outperforming any open-weight model developed in the US or Europe” and ranks among the top five globally on the Artificial Analysis Cyber Index; treat those as company claims until an independent board shows the same snapshot.
One detail Mistral leans on: several closed models score near zero on a cyber reproduction-and-patch test because they refuse the task, while Large 4 scores 82% on that specific test in Mistral’s write-up. That is a comparison of refusal behavior as much as raw skill, and Mistral says so in the cyber section of the post.
The catches before you budget around it
Three of them. First, the price you see today is the sale column; if the promotion lifts without a new announcement, the bill doubles on the numbers already on the docs page. Second, “open-weight” in the launch post is a promise with an ETA, not a file you can pull this morning, and no license text is up yet. Third, Mistral is still running reinforcement learning on the preview and says the model “continues to improve,” so any benchmark number from October 6 is a snapshot of a moving target.
What to watch next is simple: whether the sale price stays, whether October 31 holds on Hugging Face, and whether the weight license is Apache-style or a custom commercial terms sheet. Until those three land, Large 4 is a priced preview API with a countdown page, not a model you can self-host.
Frequently asked questions
How much does Mistral Large 4 cost?
On Mistral’s API docs page as of October 12, 2026, the sale price is $0.68 per million input tokens, $0.07 for cached input and $2.09 for output. The listed original prices are $1.36, $0.14 and $4.18. Mistral has not published an end date for the sale on that page.
When will Mistral Large 4 open weights be released?
Mistral’s October 6 launch post says weights drop by the end of October. Its Hugging Face placeholder lists a current ETA of October 31, 2026. The weights were not downloadable when we checked on October 12, and the license text has not been published.
How many parameters does Mistral Large 4 have?
Mistral’s docs list 1.05 trillion total parameters with 52 billion active, plus a 1.6 billion-parameter vision encoder. Hugging Face describes 49 billion active parameters per token, or 52 billion including embeddings and output layers. Context window is 1 million tokens.
Can I self-host Mistral Large 4 today?
No. The public preview is API-only on Mistral Studio. Self-hosting waits on the weight release. Until then there is no third-party host of the same checkpoint.
Sources
All accessed October 12, 2026.
- Mistral, “Introducing Mistral Large 4” (Oct 6, 2026): mistral.ai
- Mistral docs, Mistral Large 4 model page (sale and list prices, 1M context, v26.10 preview, features; checked Oct 12, 2026): docs.mistral.ai
- Hugging Face, mistralai/Mistral-Large-4.0-1T05-A52B (open-weights ETA Oct 31, 2026; 49B active per token / 52B with embeddings; checked Oct 12, 2026): huggingface.co
- Mistral AI on X, Large 4 launch thread (Oct 6, 2026, about 9:02 AM ET): x.com
Related: GPT-6.1 Sol Ultrafast costs more than Astra at standard speed · Claude Haiku 5.5 matches GPT-6 Luna’s 10-cent price · More LLM news
Last updated: October 12, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
