Europe’s best known AI lab has its biggest model yet. On 6 October Mistral released Mistral Large 4 in preview, a multimodal model with about 1 trillion parameters, of which 49 billion are active for any answer. It is live now in Mistral Studio, and Mistral says the open weights will follow at the end of October.
The price is the headline. At about $1.91 per million input tokens and $5.87 per million output tokens, it undercuts the leading American models by a wide margin.
What Mistral Large 4 is
Mistral Large 4 uses a mixture of experts design, so only a fraction of the model works on each request. That keeps it fast and cheap to run despite its size. It handles text and images, and Mistral says it was trained on data in more than 160 languages, including every official language of the European Union.
Training took place on 3,800 NVIDIA Grace Blackwell chips in Mistral’s own European data centres. That matters to Mistral’s pitch: a frontier class model built and deployable in Europe without relying on American cloud providers.
Chief executive Arthur Mensch presented the model at the AI Everything event in Abu Dhabi, and Mistral is pitching it squarely at governments and companies that want frontier performance without sending their data to the United States.
How it compares on price
Mistral bills in US dollars, so these figures use the Reserve Bank rate of 22 September. For comparison, GPT 6.1 Sol costs about $2.81 per million input tokens and $14.04 per million output tokens. Output is where most of the bill sits for long answers, and there Mistral Large 4 costs well under half as much.
Once the weights are public, businesses will also be able to run it on their own hardware or with a local provider, paying for compute rather than per token.

What Mistral claims on performance
Mistral’s own benchmarks put Large 4 ahead in several specialist areas. It says the model scored 82% on reproducing software vulnerabilities, above Claude Opus 5.5 and GPT 6 Astra, and 61.7% on the DeepSWE coding test, ahead of DeepSeek V4 Pro and Qwen 3.8 Max. It also claims wins over GPT 6 Astra on legal and finance tests. On visual grounding it reports a narrow lead over GPT 6 Astra, 42% to 41% on its Dense 200 test.
These are the company’s own numbers on its chosen tests. Independent results will matter more, and they will be easier to check once the weights are released.
Why Australian businesses should care
For Australian organisations with strict data rules, open weights are the real story. A capable model you can run inside Australia, on your own servers or a local cloud region, keeps sensitive data onshore in a way a hosted American service cannot always promise.
That is not automatic. Running a model of this size needs serious hardware, and data residency is about more than where the model sits, as our guide to AI data residency in Australia explains. But it adds a credible option for regulated sectors weighing cost against control.
Europe is back in the race
For two years the frontier has been an American story, with Chinese labs setting the pace on open models. Mistral Large 4 puts a European name back near the top, at a price that will make buyers look twice.
The test comes at the end of October, when the weights go public and anyone can check the claims. Until then, it is worth a trial on the tasks where your current model costs the most, a comparison we often run in our AI consulting projects.
