Mistral Large 4 Preview: Pricing and Planned Weights

A paid preview before the weights

On Tuesday, October 6, 2026, Mistral opened a public preview of Mistral Large 4, nicknamed Le Chonk, through its API and Mistral Studio. Its model page lists the preview at $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens, with a 1 million token context window, structured outputs, function calling and Mistral's agent tools. [1][3][2]

The weights were not part of the launch. Mistral said they would follow by the end of October, along with architecture details and its post-training method. Until then it was red-teaming the model with cybersecurity firms, vetted partners and state authorities, who got the same model with reduced moderation and expanded cyber capabilities. Mistral's VP of science, Pierre Stock, told TechCrunch the aim was to make sure the open weights could be used to defend but not to attack. [1][2]

So the model Mistral calls open-weight was, at launch, a hosted model with a promised license and release date. Mistral's model page does label it open-weight. Whether the downloadable version matches the preview is a separate question, because Mistral also said reinforcement learning was still running. [1][3]

Size and training

The model page describes a granular mixture-of-experts model with 1.05 trillion total parameters, 52 billion active, and a 1.6 billion parameter vision encoder; the launch post rounds the total to 1 trillion. It takes text and images as input. Mistral called it its largest and most capable model to date, and said more than 160 languages were in the training data. [3][1]

Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers and serves the preview from the same machines; Stock gave TechCrunch a round 4,000. The reinforcement learning run behind the preview was still in progress, producing about 33 billion tokens a day on roughly 3,000 GPUs, and Mistral said it expected large gains in the following weeks. It called the model the first milestone funded by its €3 billion Series D. [1][2]

Mistral's numbers

Mistral led with cybersecurity. It said the model ranks among the top five on the Artificial Analysis Cyber Index, scores 82 percent on one of that index's tests, which asks a model to reproduce a real vulnerability and then patch it, and solves 93 percent of Cybench. It added that Claude Opus 5.5 and GPT-6 Astra score near zero on the reproduce-and-patch test because they refuse the task, and that the model's refusal rate on malicious cyber prompts was higher than any other open model's. [1]

On coding it reported 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA and 28.3 percent on Terminal-Bench 4, and a blind human evaluation run with Surge AI in which it placed second of five, behind Claude Opus 5 and ahead of Kimi K3, GLM 5.3 and GLM 5.2. It also claimed wins over GPT-6 Astra on a visual grounding test (42 to 41 percent) and on vals.ai finance and legal tasks. [1]

These are Mistral's selections, some run by third parties it chose and some internal. The cyber comparison measures a refusal as a failure, which says as much about the closed models' safeguards as about capability. TechCrunch described the benchmark results as still pending on launch day. [1][2]