✳ the wire · analysis

Mistral Large 4 is out in public preview: 1T parameters, $0.68 in and $2.09 out for now, open weights by the end of October

Mistral Large 4confirmedby ArtificialWatch

Get launch alerts like this, free →

Mistral Large 4 is out in public preview: 1T parameters, $0.68 in and $2.09 out for now, open weights by the end of October
Source imagery · verified against a primary source

Mistral launched Mistral Large 4 today as a public preview on its own API. Per Mistral's docs it is a natively multimodal mixture-of-experts model with about 1.05 trillion total parameters and 49 billion active, plus a 1.6B vision encoder. Mistral says it will release the open weights by the end of this month.

THE PRICE

Right now it costs $0.68 per million input tokens, $0.07 for cached input and $2.09 for output. Mistral's docs show those prices next to a struck-through list price of $1.36, $0.14 and $4.18, and OpenRouter charges the lower rate too. Mistral does not say how long the lower price lasts. Mistral's docs give a 1M-token context window; OpenRouter lists 512K.

Correction: an earlier version of this story gave the $1.36 / $4.18 list price as the current price.

WHAT MISTRAL SAYS IT SCORES

These are Mistral's own figures, several taken from Artificial Analysis's coding index:

- DeepSWE v1.1: 61.7% - SWE-Atlas QnA: 59.4% - Terminal-Bench 4: 28.3% - Coding Agent Index: 49.8%, which Mistral says puts it ahead of DeepSeek V4 Pro and Qwen 3.8 Max - Cybench: 93%, and top five on the Artificial Analysis Cyber Index, per Mistral

Independent: Artificial Analysis scores the preview 38 on its Intelligence Index, 64th of 225 models, at about 116 output tokens per second.

In a blind coding review Mistral ran with Surge AI, the preview placed second of five models, behind only Claude Opus 5. Mistral calls it the strongest open-weight model built in the US or Europe.

HOW IT WAS BUILT

Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, which also serve the preview. Mistral says the reinforcement learning run is still going and the model will keep improving, and that the training data spans more than 160 languages.

WHAT WE DO NOT KNOW

The weights license and the exact release day ("end of this month" is Mistral's window). Until then, Mistral says, cybersecurity partners and state authorities are red-teaming a version with reduced moderation. Our launch alert fired at 13:37 UTC, when OpenRouter listed the model as mistralai/mistral-large-4-0.

Source: Mistral ↗ · Mistral Large 4 tracker · the bench index

sweeping every 60 seconds

Know the minute it drops — not the minute we write it up.

Claude Opus 5 went live at 16:51 UTC. The alert was in subscribers' inboxes at 16:52.

  • Free forever
  • No card
  • Unsubscribe in one click

← back to the wire