Mistral Large 4 self-hosting vs API: the math and the break-even point
Mistral Large 4 is free to download, but the server costs about $38,534 a month. How many daily tasks make the API cheaper.
Mistral Large 4 self-hosting costs almost $38,534 a month if you rent eight B200 GPUs, even though the model is free to download. With an assumption of 50,000 input tokens and 5,000 output tokens per task, the break-even point against Mistral's API is about 28,000 tasks a day. Below that, the API is cheaper. Here is the breakdown from the video, in article form.
What changed: Mistral Large 4 open weights
The headline misleads. Mistral Large 4 went into public preview on October 6, 2026, and since then you could pay to use it through the API. What's new is that Mistral published the model's weights under an open license, so now you can download it and run it yourself.
Open doesn't mean free for everything. The license is a modified MIT: you can use it, modify it, fine-tune it and charge for what you build, with one exception. If you are going to sell access to the model as a service, you need a separate commercial agreement. And the full technical report, according to Mistral, comes later.
How it works under the hood
Mistral Large 4 has about one trillion parameters, but it is a mixture-of-experts model: about 52 billion work at each step. That makes it fast to run, but not light, because all the parameters have to be loaded into memory. That is why the official example uses eight GPUs at once.

It also sees images, reads up to one million tokens of context and has three reasoning modes: low, medium and high, which is the default.
The two paths: the API or your own server
Path one, the API. You open Mistral Studio, create a key, pick Large 4 and send your first request like with any other model. It supports function calling and structured outputs.
Path two, your own server. You pull the vLLM container Mistral prepared, start it with eight GPUs and turn on the reasoning and tool parsers. The first time, people get stuck on the four-bit quantized version (NVFP4), which targets Blackwell GPUs: if you rent others, you need a different format. And on Hugging Face, no provider offers it ready to use yet.
The math in dollars
Our assumption: an agent task with 50,000 input tokens and 5,000 output tokens, no caching. On Mistral's API input costs $0.70 per million tokens and output $2.10, which is about 4.5 cents per task. With Claude Opus 5.5, at $4 input and $20 output per million, it is 30 cents per task.
At a thousand tasks a day for thirty days, that is $1,365 with Mistral and $9,000 with Opus: a difference of $7,635 a month. That is list price, not quality. For context, AI.cc reported that open models went from 11% to 38% of its enterprise customers' tokens between the first quarter of 2025 and the first quarter of 2026; that is one provider's data about its own users.
Three business cases (our analysis)
- A software agency that reviews code daily: through the API, about 4.5 cents per task against 30 with Opus; at a thousand tasks a day, over $7,000 a month of difference.
- A company with 40,000 tasks a day: through the API that would be about $54,600 a month; renting the eight GPUs costs $38,534. Here your own server wins and saves about $16,000, not counting the engineer who keeps it running. Also, your customers' data never leaves your infrastructure.
- A team already working on Microsoft Azure: Mistral lists Large 4 inside Microsoft Foundry, so it is called from the same subscription and billing. Foundry pricing was not verified: confirm the rate before committing budget.
Who it's not for
It isn't for anyone who wants to run it on a laptop: even quantized to four bits it weighs hundreds of gigabytes. Nor for anyone who wants to resell access as a service without a commercial agreement, or who needs the very best with no trade-offs: in the blind human preference test Mistral publishes, only Claude Opus 5.5 ranked higher.
The objection: it talks twice as much
Artificial Analysis, an independent firm, measured the preview version and found it generated about 200 million output tokens against a median of 81 million for comparable models, about 2.5 times more. Applied to our task, output would rise to about 12,300 tokens and the cost to about 6 cents, still almost five times cheaper than Opus under the same assumption. For the objection to win, Mistral would have to talk about 25 times more. The sensible move is to measure the tokens of your own tasks.
What they're not telling you
The phrase going around is "the best open model outside China", but almost all the numbers on Mistral's page (19.7 on FrontierCode, 41 on Terminal-Bench, 96.6% refusal of malicious requests) are published by the maker itself. What's independent is more modest: Artificial Analysis gave it 38 on its intelligence index on October 6, eighth among open models, behind seven Chinese models, using the preview version. No new independent measurement of the final weights was found. Treat it as a candidate to test, not a verdict.
What breaks in practice
- The default mode is high: for simple tasks it spends reasoning you don't need. Switch to low and measure.
- The price lives on a page that says public preview, so it can change.
- The technical report isn't out yet: you don't know what data it was trained on or how it was evaluated.
- Without the eight GPUs, the option to download the weights doesn't exist for you.
How to apply it in your business?
The break-even point: eight B200s at $6.69 an hour each add up to $53.52 an hour, about $38,534 a month. Divided by 4.5 cents per task, that is about 847,000 tasks a month, around 28,000 a day. Five rules:
- Before deciding, run 50 of your own tasks and measure the real tokens.
- Start in low mode and go up only if it fails.
- Stay on the API until about 28,000 tasks a day.
- If you are going to sell access to the model, read the license with your lawyer.
- Compare against the model you already use, on your tasks, not on anyone's table.
The math uses a token assumption and doesn't include engineers: change either one and the result changes. For the launch context, read the full breakdown of Mistral Large 4.

Frequently asked questions
Is Mistral Large 4 free?
The weights are free to download, but running it on your own server with eight rented B200 GPUs costs about $38,534 a month. The API charges $0.70 per million input tokens and $2.10 for output.
When does self-hosting make sense?
According to our math, from about 28,000 tasks a day with the assumption of 50,000 input tokens and 5,000 output tokens. Below that, the API.
Can I sell services built on Mistral Large 4?
You can charge for what you build, but if you sell access to the model as a service you need a separate commercial agreement.
Does it run on a laptop?
No. Even quantized to four bits it weighs hundreds of gigabytes and the official example uses eight GPUs.
Is it the best open model?
That is the maker's claim. The independent measurement of the preview version ranked it eighth among open models.
Conclusion
Mistral Large 4 self-hosting doesn't save you money on its own: it gives you an option, and the math decides whether it pays off. The real question is how many tokens each of your tasks uses. Measure it, start with the API and move to your own server only if the volume justifies it. Sources: Mistral announcement, model documentation, Claude pricing and Lambda pricing.