Mistral Large 4: Frontier-Class, Multilingual, and Open Weights by Month's End

Mistral Large 4 launched on October 6, 2026, with API pricing of $1.36 per million input tokens and $4.18 per million output tokens and open weights promised for the end of October. For companies that have been waiting for a frontier-class model they can eventually run inside their own boundary, that combination is the news.

The model is a natively multimodal mixture-of-experts system with 1 trillion total parameters and 49 billion active per token. Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacenters, and says it is fluent in more than 160 languages. A public preview API is open now through Mistral Studio.

The benchmark claims are Mistral’s own, so treat them as a starting hypothesis. It reports 61.7 percent on DeepSWE v1.1 and 28.3 percent on Terminal-Bench 4 for coding, 59.9 percent on AutomationBench for agentic business tasks, and a top-five ranking on the Artificial Analysis Cyber Index. In vision, it claims 42 percent against GPT-6-Astra’s 41 percent on Dense 200 grounding — a one-point gap that is not a reason to switch models.

The more durable signal is where the model comes from. Mistral is red-teaming Large 4 with cybersecurity partners and state authorities before the weights ship. For organizations with data-residency rules, multilingual customers, or regulators who ask where inference runs, a lower-priced frontier option with open weights changes the shortlist.

What to do with it: when the weights land, run your own evaluation set on your own documents, in the languages your customers actually use, and price the result against whatever model you run today. A vendor benchmark tells you where to look. Your workload tells you whether to ship.