You work for yourself, you already pay for one AI subscription (maybe two), and now a French model turns up promising to rival the best. Should you change everything?
Mistral Large 4 is the new model from Mistral AI, announced on 6 October 2026. It ships as a preview (a trial version that Mistral is still improving), and it has 1 trillion parameters. Hence its in-house nickname: "le Chonk", the word the internet keeps for very round cats.
So let's look at what it can do, what it costs, and what it changes for you.
What is Mistral Large 4?
It is the largest model Mistral has ever released, and the first since its €3 billion funding round.
A parameter is a setting the model learnt during training. The more it has, the more it can remember. Mistral Large 4 has 1 trillion of them, but it only uses 49 billion at a time.
It is a bit like a hospital: it employs hundreds of specialists, but for your knee you only see two or three. The model does the same and only wakes the "specialists" your question needs. Each answer therefore takes far less computing than its size suggests.
According to Mistral's announcement, here is what it can do:
- Read text and images. A contract in PDF, a chart, a photo of a floor plan. It answers in text;
- Work in more than 160 languages, including every official language of the European Union. Handy if your clients are not all in one country;
- Chain tasks on its own. Look something up, fill in a spreadsheet, produce a finished document.
How much does Mistral Large 4 cost?
Mistral lists $1.36 per million input tokens, and $4.18 per million output tokens.
A token is a piece of a word. Input is what you send the model (your request, your documents). Output is what it writes back.
In practice? Imagine you have it draft a reply to a client's email, so about 2,000 tokens to read and 1,000 to write. That costs you less than one cent. For a thousand emails, expect roughly $7.
Here is the same job with two models released in recent weeks:
| Model | Input, per million tokens | Output, per million tokens | A thousand email replies |
|---|---|---|---|
| Mistral Large 4 | $1.36 | $4.18 | About $7 |
| GPT-6 Sol | $2 | $10 | $14 |
| Claude Sonnet 5.5 | $2 | $10 | $14 |
So the bill is cut in half! We compared the other two in our article on GPT-6 Sol and Claude Sonnet 5.5.
This is the price of the API, the access a program or an automation uses to talk to the model. You pay for what you use, with no subscription. The announcement says nothing about Le Chat, Mistral's assistant for the general public.
How can you try Mistral Large 4 today?
At the time of the announcement there was only one way in, the API preview in Mistral Studio (Mistral's space for people who plug its models into their tools).
If you already have an automation that calls an AI model, in n8n or Make for example, trying it is quick. You change the model at the step in question, then compare the result. Our n8n course shows how to plug a different model into each step.
If you only use an assistant such as ChatGPT or Claude to write and think, you have nothing to do for the moment.
A preview moves. Mistral says the model is still being trained, so its answers may change from one week to the next. Try it alongside your usual tool, and keep that one for the work you deliver to a client tomorrow morning. 😉
What does "open weights" mean?
Mistral promises to publish the model's weights before the end of October. The weights are the parameters we talked about above: the model itself, as files anyone can download.
Imagine a great restaurant publishing the exact recipe for its best dish. Anyone can read it, but you need a restaurant kitchen to pull it off. It is the same here. With 1 trillion parameters, le Chonk will not run on your computer. It needs whole servers.
So what do you gain? When a model is open, hosts other than its maker can offer it. You no longer depend on a single provider, or on a single price.
Mistral has not yet said, though, under which licence the weights will be released. Yet the licence decides what you are allowed to do with them, commercial use included. Answer at the end of the month!
Until then, Mistral is having the model tested by cybersecurity specialists and public authorities, to find out what someone with bad intentions could make it do.
Does your data stay in Europe?
I have lost count of the times I hesitated before pasting a client's contract into an AI (and ended up anonymising the whole thing by hand...).
Mistral trained its model on 3,800 NVIDIA graphics processors, in its own data centres in Europe. The preview runs on those same machines.
The company also announces a European version that it operates alone, without going through another provider and under European law. If your clients ask where their documents go, you will have a simple answer for them.
Hosting in Europe does not let you off anything. It is still up to you to know what your client allows you to send to a tool, whichever tool it is.
Is Mistral Large 4 better than ChatGPT or Claude?
Read Mistral's promise carefully: its model beats every open model built in the United States or Europe, and it rivals the best open models in the world. The word that counts is "open".
Against the closed models from OpenAI and Anthropic, it is less clear-cut:
- On office work, it completes 59.9% of the 657 tasks in a test run inside Gmail, Google Sheets, Slack and Salesforce. So it fails four in ten, and you will have to check its work;
- On law and finance, Mistral says it beats GPT-6 Astra, OpenAI's top model, in tests handed to an outside evaluator;
- On code, it stays behind. It scores 61.7% on the DeepSWE v1.1 test, where OpenAI reports 68.8% for GPT-6 Sol. And when professionals rated its code without knowing which model had written it, it came second out of five, behind a Claude model.
These figures come from the vendors themselves, and each one picks the tests that suit it. The French site Abondance points this out too: no complete independent evaluation of this preview has been published yet.
Where Mistral pushes hardest is cybersecurity, where its model ranks among the top five in the world on an independent test. Unless you sell security audits, that will not change your days.
Should you switch tools now?
It all depends on how you use AI today:
- You chat with an assistant to write, summarise or think. Change nothing, Mistral Large 4 was not announced in any assistant at launch;
- You have automations working for you, and an API bill that keeps climbing. Try it on one repetitive task, then compare the result and the price;
- You handle sensitive documents for your clients. Wait for the weights and the licence, at the end of October, before you decide.
Does one more model change anything for your business? Your clients pay you for a result, and they will never ask which model produced it.
We will know more at the end of October, once the licence is out. 😉
Frequently asked questions about Mistral Large 4
Is Mistral Large 4 free?
No. Its API costs $1.36 per million input tokens and $4.18 per million output tokens, with no subscription.
Its weights will be published at the end of October 2026. Downloading them is one thing, running them is another. You need servers, so a host to pay.
Is Mistral Large 4 available in Le Chat?
The announcement of 6 October 2026 does not mention it. The model is only offered through the API, as a preview, in Mistral Studio.
If you use Le Chat to write or summarise, nothing changes for you today.
When will the Mistral Large 4 weights be released?
Mistral announces a release before the end of October 2026, once its safety testing is finished.
The licence is not known yet. It will say whether commercial use is allowed.
Can you install Mistral Large 4 on your own computer?
No. A model with 1 trillion parameters needs whole servers, not a laptop.
For someone working alone, open weights mainly let you go through a host other than Mistral, if the licence allows it.
Is Mistral Large 4 better than GPT-6 or Claude?
Not everywhere. Mistral places it ahead of every American and European open model, and ahead of GPT-6 Astra on law and finance tests.
On code it is behind, with 61.7% on the DeepSWE v1.1 test against the 68.8% OpenAI reports for GPT-6 Sol. These results come from the vendors and have not yet been confirmed by a complete independent evaluation.