Mistral Large 4: we tested Europe's big open AI model on launch day
Mistral Large 4 is a huge open AI model from France, with data stored in the EU by default. We tested it on launch day: cheap and nearly perfect, with one small slip.
Mistral is a French AI company. On 6 October 2026 it launched Mistral Large 4, its biggest model so far. Mistral says it beats every open model made in Europe or the US. You can use it online now, and you will be able to download it at the end of October. We tested it on the day it came out. It did our coding tasks well and very cheaply: six tries cost 1.1 cents in total. But it made one small mistake that no other model in our tests made.
What is Mistral Large 4?
Mistral Large 4 is a "public preview". That means it is not finished, and Mistral can still change it. You can use it through Mistral's API now. An API is the connection that programs use to talk to an AI.
The model is huge: about a trillion parameters. Parameters are the numbers a model learns during training. But for each word, only about 50 billion of them do the work, which makes it cheaper to run. It can read text and images, and up to one million tokens at once.
Mistral says it trained the model in its own data centres in Europe. The "weights", the files you need to run the model yourself, are coming at the end of October, says its announcement.
What does it cost?
On its model page, Mistral shows two prices for Large 4. The old price is struck through, and a lower price is next to it:
- Reading: $0.68 per million tokens (struck through: $1.36).
- Writing: $2.09 per million tokens (struck through: $4.18).
The announcement itself still lists the higher price. Mistral does not say why there are two prices or how long the lower one will last. We paid the lower price through OpenRouter, a service that gives access to many AI models.
At the lower price, Mistral Large 4 is cheaper to write with than Gemini 3.8 Flash, DeepSeek V4 Pro, Grok 4.7, GPT-6.1 Sol and Claude Sonnet 5.5. Only DeepSeek's small V4.1 Flash is cheaper.
Our test: cheap and nearly perfect
We gave Mistral Large 4 the same two coding tasks as in our DeepSeek V4 test. In the first, it had to fix a bug in code that reads times like "1h30m". In the second, it had to build a Dutch invoice calculation from a description. It did each task three times. We checked every answer with our secret tests: tests we wrote ourselves and only ran after the AI had finished.
- Bug fix: 23/24 every time. That is the same score as every model we have tested. The one test it missed checks a rule our task did not mention.
- Invoice: 14/15, 14/15, 15/15. So it missed one test in two of its three tries.
It was also very short and to the point. Over six tries it wrote about 4,000 tokens. DeepSeek V4 Pro wrote about 25,000 for the same work, because it "thinks" at length before it answers. So Mistral cost 1.1 cents, against 2.3 cents for DeepSeek V4 Pro.
One more thing: on launch day, 3 of our first requests failed with an error from the provider. We ran them again a few minutes later, and then they worked.
Want to see the test? One secret test checks that "1h banana" is refused, because there is junk after the time. Another checks "1h1h", the same unit twice: every model so far has let that one through. Our test page shows all three jobs, the secret tests and the exact words we sent. You can also download them and run them on any AI yourself.
The one test it missed
The test Mistral missed is a small but real trap. An invoice line has a quantity of 1.005, for example 1.005 kilos. That is allowed: our task said up to three decimals.
Mistral's code checked the decimals by multiplying by 1000. But a computer stores 1.005 a tiny bit wrong, so it gets 1004.9999999999999 instead of 1005. Mistral's code then said "too many decimals" and refused the line.
This kind of mistake is easy to miss. It only shows up with some numbers. Every other model we have tested, from GPT-6.1 Sol and Claude to Grok and DeepSeek, passed this test every time.
Where does your data go?
This is where Mistral is different from DeepSeek. "By default, your data is hosted in the European Union", Mistral's help center says. Only if you choose its US connection is your data stored in the United States. For some features, data can go outside the EU for a while, to other companies that Mistral uses.
Mistral may use what you send to train its models. You can turn this off, the help page on training explains. The setting for the chat app and the setting for the API are separate, so you have to switch off both.
Mistral or DeepSeek?
Both are big open models at a low price. The main difference is where your data goes when you use the company's own service:
- Mistral: in the EU by default, under European privacy law.
- DeepSeek: in China, says its privacy policy.
On our coding tasks, DeepSeek V4 Pro passed every secret test. Mistral Large 4 made one slip, but cost less than half as much. For work with private or customer data in Europe, Mistral is the easier choice. Just check the code it writes, like you would with any AI.
Want to compare with a paid tool? In our Claude Code test, one coding task cost between 14 and 54 cents.
How we checked
On 6 October 2026 we read Mistral's announcement, its model page and two pages from its help center about where data is stored and training. The screenshots are our own captures of those pages, cropped only. The prices of other models come from their makers' own pages, read between 4 and 6 October.
We ran the test on launch day through OpenRouter, with Mistral itself answering every request. We used default settings, three tries per task, and the same secret tests as in our earlier tests. The owner of this site paid for the requests. The charts were made by us. None of the images was made by AI.
This was a small test with two clear tasks, on the first day of a preview. The model may still change before the weights come out. Prices can change too, so check Mistral's page before you build on it.