Mistral Large 4 opens a 1-trillion-parameter preview API at $1.36 per million input tokens

Mistral calls the model ML4 and, less officially, le Chonk. It is the company's largest model so far, a mixture-of-experts design that combines instruction following and reasoning in one network. You can call the preview API on Mistral Studio now; what you cannot do yet is download it, because the open-weight release is scheduled for the end of the month after a red-teaming phase.

Size, training, and price

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe, and the preview runs on the same hardware. A large share of the training data was multilingual, across more than 160 languages including every official language of the European Union.

Pricing on the product page is $1.36 per million input tokens and $4.18 per million output tokens. Mistral says the model will be available in several regions, including a European deployment it operates end to end under European law.

Two things the announcement does not state are the context window and the license the weights will carry. Mistral says architecture details, more benchmarks, and the post-training method follow with the weights.

What Mistral's benchmarks claim

According to the company's numbers, ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, for a Coding Agent Index of 49.8%. In a blind human evaluation run with Surge AI, annotators rated it 3.74 on a five-point scale, second of five models and behind only Claude Opus 5 at 4.22.

On cybersecurity Mistral reports 93% of the 40 Cybench challenges solved and 82% on the Artificial Analysis Cyber Index task that reproduces and patches a real vulnerability. The company adds that Claude Opus 5.5 and GPT-6 Astra score near zero on Cybench because they refuse the task, and that ML4 nonetheless refuses malicious cyber prompts more often than other open models.

On visual grounding, Mistral puts ML4 at 42% on Dense 200 against 41% for GPT-6 Astra, and on Lakera's B3 security benchmark it reports 93.3% of prompt-injection attacks resisted. None of these figures has an independent check yet.

Why the weights matter more than the preview

Until the weights ship, ML4 is an API product like any closed model, priced in the same band. The company frames the open release as the point: security teams that hit refusals in closed models, or that cannot send data outside their own infrastructure, get a model they can run on their own hardware under their own policy.

Mistral ties the model to its €3 billion Series D and says the reinforcement-learning run behind the preview is still going, so the preview is a snapshot rather than a final version.