Mistral Large 4 went into public preview on October 6, 2026, at $1.36 per million input tokens and $4.18 per million output tokens, the prices on Mistral's launch post. The API is live. The weights aren't, and the license hasn't been named. Mistral says it will release the weights by the end of the month, so any plan that depends on self-hosting is a plan against a promise, not a product.
What Mistral launched on October 6
According to Mistral's launch post, Large 4 is a "1 trillion-parameter natively multimodal model with 49 billion active parameters." The preview runs through Mistral Studio. The post also says architecture details and additional benchmarks will follow later, and that the model covers 160+ languages, including every official EU language.
Mistral's own pages disagree on size. The post says 49B active parameters. The docs model card says 52B active, 1.05T total. Both are Mistral's, and neither is marked as a correction, so treat "about 1 trillion total, roughly 50B active" as the safe summary. The model card lists a 1M-token context window.
The model ID is also shown two ways. The card's URL and the models overview use mistral-large-4-0, while the card page itself displays mistral-large-4 with a "+1" alias. Copy the string from Studio rather than from this article.
Mistral's docs call the model "open-weight" and list its license as "Open." That's a label, not a license. The launch post names none, so don't read "open source" into it, and don't assume the terms match Large 3's. Large 3's model card lists Apache 2.0, but nothing Mistral has published says Large 4 will follow.
Mistral Large 4 pricing vs GPT-6 and Claude
All figures below are per million tokens, from each vendor's own docs.
| Model | Input | Output | Context |
|---|---|---|---|
| Mistral Large 4 | $1.36 | $4.18 | 1M |
| Mistral Large 3 | $0.50 | $1.50 | 256k |
| GPT-6 Astra | $10 | $50 | 1,050,000 |
| GPT-6 Sol | $2 | $10 | 1,050,000 |
| Claude Opus 5.5 | $4 | $20 | 1M |
| Claude Sonnet 5.5 | $2 | $10 | 1M |
Large 4 prices come from the model card, and the $1.36/$4.18 pair also appears on the launch post. The card lists $0.14 for cached input. The model card shows a second set beside these, $0.68 input, $0.07 cached and $2.09 output. Mistral's page doesn't say in text what the second set is, so this article compares on the $1.36/$4.18 pair shown on the launch post. If $0.68/$2.09 is the price you pay, every Mistral ratio in this section roughly halves. Check the card before you budget.
The other rows come from OpenAI's GPT-6 Astra page, the GPT-6 Sol page, Anthropic's Opus 5.5 overview and its Sonnet 5.5 overview. Astra also lists $1 cached input. Astra and Sol list 1,050,000 tokens of context; the Sol page also says requests above 272K input tokens are billed at 2x.
On paper, Large 4's output price is less than half of Sonnet 5.5's and about 8% of Astra's. Against its own predecessor it's a step up, roughly 2.7x on input and 2.8x on output. Those are price ratios and nothing more. Those ratios use the $1.36/$4.18 pair; at $0.68/$2.09 they'd be about half. The models haven't been benchmarked on cost per task, so "cheaper" tells you nothing about what a finished job costs. For the OpenAI and Anthropic side of the comparison, see our GPT-6 Astra pricing breakdown and the Claude Opus 5.5 re-pricing guide.
Where the data is processed
This is the part to read slowly. Mistral says Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that the public preview "is served on that same infrastructure." The post also describes "a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law," but presents it as part of availability across multiple regions that is still coming.
Mistral's regional inference documentation lists three endpoints. api.mistral.ai is global and "does not commit to a specific inference location." api.eu.mistral.ai uses EU and EFTA data centres, and api.us.mistral.ai the US. Regional endpoints bill at 1.1x the base price per Mistral's docs. Two more caveats from the same page: the control plane (account configuration, API keys, billing, usage analytics) isn't regional, and regional endpoints only serve models hosted in that region.
We haven't verified that mistral-large-4 is served on api.eu.mistral.ai today. Test the endpoint with a real request and ask Mistral in writing. Mistral's docs say available models vary by region and recommend checking before routing production traffic. None of the Mistral pages we read for this article, the launch post, the model card and the regional-inference documentation, makes a GDPR compliance claim for Large 4, and this article doesn't either. If a contractor or vendor will touch the data around the model, who can access that data under GDPR is a separate question that deserves its own answer.
What the open weights would change
If the weights ship, the post says that for security operations it "will be able to run on private cloud or on-premise." That's the real draw for a company that can't send data to a US API or even to a European one.
Three things are missing, and all three decide whether it's practical. The license text isn't out, so commercial-use terms and any user-count limits are unknown. The release date is only "by the end of the month." And Mistral hasn't published serving requirements. The only hardware figure anywhere is for training, 3,800 GPUs, which says nothing about what inference takes for a model this size.
For contrast, DeepSeek's open-weight release came with an MIT license and a stated parameter count, as we covered in our DeepSeek V4.1 Flash piece. Mistral's is a promise so far.
What Mistral says about performance
All figures below are taken from Mistral's launch post. We haven't checked them against the evaluators' own sites.
The post attributes 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4.0, 59.9% on AutomationBench and a 49.8% Coding Agent Index score to Artificial Analysis; the Index puts Large 4 "ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max." Mistral cites Artificial Analysis's Cyber Index as ranking Large 4 among the top five models globally, 93% on Cybench from its own internal testing, and 93.3% on Lakera's B3.
In a blind evaluation by Surge AI, in which professional annotators rated coding outputs on a 1-5 scale, Large 4 scored 3.74, second of five, behind Claude Opus 5 at 4.22. It's the one figure here that shows a gap to a named leader.
Mistral also claims Claude Opus 5.5 and GPT-6 Astra "score near zero" on one cyber test because they refuse the task. That's a vendor describing its competitors' results. Treat it as Mistral's claim, and run your own evals.
What this means for your AI and hosting decision
If you handle EU personal or confidential data, you have three options: stay on a US API, use Mistral's EU endpoint, or plan to self-host. Only the first two exist today, and the second comes with an unverified model-availability question.
Our read, and it's judgment, not a sourced fact: run a small evaluation against the preview now, using api.eu.mistral.ai if it serves the model. Budget the 1.1x regional uplift. Hold any self-hosting spend until the license text and serving requirements are published. And don't migrate production on benchmarks Mistral wrote about itself, from a model still in preview.
On cost, the launch-post price of $1.36/$4.18 is below the $2/$10 to $10/$50 you'd pay for the GPT-6 and Claude rows above, but you can only call that a saving after your own task-level test. For where an AI line fits in the wider plan, see how to budget AI seats and infrastructure at Series A. For other releases and price changes, there's AI model releases and pricing: what changes for engineering teams.
FAQ
Is Mistral Large 4 open source?
Not on what's published. Mistral's docs call it "open-weight" and list the license as "Open," but the launch post names no license, and the weights aren't released yet. Open weights and open source are different things, and the license text will decide what you can do with them.
Does Mistral Large 4 keep data in the EU?
Mistral says the preview is served on its European infrastructure, and its docs describe an EU and EFTA endpoint at api.eu.mistral.ai, billed at 1.1x. We couldn't verify that Large 4 is served on that endpoint today. The global endpoint makes no location commitment, and the control plane isn't regional. Test it and ask Mistral.
How much does Mistral Large 4 cost compared to GPT-6 and Claude?
The launch-post price is $1.36 input and $4.18 output per million tokens. GPT-6 Astra is $10 and $50, GPT-6 Sol and Claude Sonnet 5.5 are $2 and $10, and Claude Opus 5.5 is $4 and $20. These are per-token prices, not cost per task.
When are the weights released and what hardware is needed?
Mistral says "by the end of the month," so by the end of October 2026. It hasn't published serving requirements, so there's no reliable hardware figure yet.
What is the model ID and context window?
The docs use mistral-large-4-0 in the URL and overview, and the card page shows mistral-large-4. Copy the ID from Mistral Studio. The model card lists a 1M-token context window.