Mistral Large 4

Mistral Large 4 is the largest model Mistral has built and the first to put a trillion parameters behind a European price card. Announced October 6, 2026 as a public preview on Mistral Studio and the Mistral API under the name mistral-large-4-0, it is a mixture-of-experts with roughly 1 trillion total parameters and 52 billion active per token, natively multimodal. The price has two numbers and you should know which one you are quoting: Mistral's own pricing page lists $1.36 per million input tokens, $0.14 cached, and $4.18 output as the original price, and is currently charging a sale rate of $0.68, $0.07, and $2.09, with no end date published for the sale. Budget the list price and enjoy the discount while it lasts. Mistral's model card advertises a 1 million token context window, though third-party registries showed 512K on launch day and OpenRouter documents a 262,144 token completion cap, so verify the window your provider actually serves before you design around it. The launch post publishes a full self-reported table: DeepSWE v1.1 at 61.7 percent, SWE-Atlas-QnA at 59.4, Terminal-Bench 4 at 28.3, a Coding Agent Index of 49.8, AutomationBench at 59.9, 1,393 Elo on AA-Briefcase, Cybench at 93 percent, 82 percent on a vulnerability reproduce-and-patch test that Mistral says is the highest reported, and 42 percent on Dense 200 visual grounding against 41 for GPT-6 Astra. The most honest row is the blind human coding evaluation Mistral commissioned from Surge AI, where Large 4 scored 3.74 out of 5, second of five and behind Claude Opus 5 at 4.22. The open-weight part is a promise, not a fact yet: Mistral said weights drop at the end of October 2026 and has named no license, and it says it is red-teaming the model with cybersecurity leaders, vetted partners, and state authorities until they ship. Treat this as a hosted preview with an open-weight roadmap, and note that Mistral still calls Medium 3.5 its frontier-class agentic and coding model.