Mistral AI released a research public preview of its Large 4 model this week. The model has about one trillion parameters, of which 49 billion are active at any time. Independent benchmark firm Artificial Analysis (AA) gave it a score of 38 on its Intelligence Index, the highest for any model coming from outside the United States or China.
How it performed in tests
AA’s Cyber Index gave Large 4 a 50‑point score, matching the Chinese GLM‑5.3‑Flash and trailing only the higher‑scoring MiMo‑V2.6‑Pro (56). Its best result was on the CyberGym‑E2E‑AA cybersecurity task, where it scored 82 %, ahead of MiMo‑V2.6‑Pro (79 %) and the max‑score GPT‑6 Luna (78 %). The model also improved document and image reasoning, reaching 19 % on the GDP.pdf test, the same as MiMo‑V2.6‑Pro.
Price and availability
Mistral charges $1.36 per million input tokens and $4.18 per million output tokens, with a $0.14 fee for cached input. That works out to $1.13 per Intelligence Index task. For the first two weeks the preview is offered at a 50 % discount, lowering the cost per task to $0.57. The company says the full model weights will be released at the end of October 2026.
Why it matters
For developers who need large‑scale language models, Mistral Large 4 offers a high‑capacity option that outperforms other non‑US/China open models on several benchmarks. However, its higher price means it may be less attractive for cost‑sensitive projects until pricing drops or competition improves.
