Mistral's 1T Open-Weight 'Le Chonk' Slashes Frontier Output Pricing ~24x — and Hands Closed Labs a $0.68 Problem

Mistral's own benchmark dossier positions [Mistral Large 4] in the same league as the most capable closed models — and ahead of every open-weight model developed outside China on cybersecurity. On the Artificial Analysis Cyber Index it ranks among the top five globally; on a curated vulnerability-reproduction and patch test it scores 82%, the highest of any model tested; on Cybench (a 40-exercise agentic-cyber suite) it solves 93%. On Lakera's B3 AI security benchmark it resists 93.3% of attacks with no higher peer score recorded. The combination of code and agentic scores —61.7% on DeepSWE v1.1, 49.8% on the Combined Coding Agent Index, 1,393 Elo on AA-Briefcase — puts it in striking range of [Anthropic]'s Claude Opus 5 family, which still wins the blind human evaluation 4.22 to 3.74.

Mistral Large 4 is priced at $0.68 input / $2.09 output per million tokens on Mistral's hosted API — a 36% premium over Mistral Large 3 ($0.50/$1.50) but a fraction of closed-frontier pricing. Output tokens are the variable that actually bills enterprise workloads; on that line, ML4 is roughly 24x cheaper than [OpenAI]'s GPT-6 Astra ($50/MTok output), ~5x cheaper than [Anthropic]'s Claude Opus 5.5 ($20/MTok) and Claude Opus 5 ($25/MTok), and ~5x cheaper than [OpenAI]'s freshly launched GPT-6.1 Sol ($10/MTok) introduced at OpenAI DevDay on September 29, 2026. Even DeepSeek — the Chinese open-weight yardstick that reset the floor in 2025 — lists V3.2 at $0.42/MTok output and V4 Pro at $0.87/MTok, both now undercut by [Mistral]'s output rate only when cached-input economics are ignored.

Open-weight release without capital is a hobby project. Mistral's combination — Apache 2.0 weights, [NVIDIA] Grace Blackwell training silicon, [Microsoft] Azure distribution, and [Samsung Electronics] foundry backing — turns [Mistral Large 4] into infrastructure for the European sovereign-AI agenda. Mistral Medium 3.5 already ships in Microsoft Foundry and Foundry Local; Microsoft committed in July 2026 to fund Mistral's European data-center build-out with thousands of [NVIDIA] Vera Rubin GPUs. [Samsung Electronics] led the €3B Series D on September 8, 2026 and signed a parallel strategic partnership to embed [Mistral] inside its semiconductor fabs for defect detection and yield optimization on advanced memory and logic chips.

[Mistral Large 4] is not the only1T-class open-weight model — and not the largest. Moonshot AI's Kimi K2.5 (January 2026) and K2.6 (May 2026) both crossed 1T total parameters with 32B active per token and a 256K context window; Moonshot's Kimi K3 reached 2.8T parameters in July 2026, the largest open-weight LLM on record at a 57 Artificial Analysis Index score. Z.ai's GLM 5.3 (listed alongside [Mistral] on Mistral's own models page) ships with a 1M-token context window. DeepSeek V3.2 and the newer V4 Pro hold the cheapest frontier-class inference prices globally. The novelty of [Mistral Large 4] is not scale — it is the convergence of European capital (Samsung), Western hyperscaler distribution (Azure + Bedrock), and frontier-grade benchmarks inside a sovereign-AI package.

The instinct to read \"open weights\" as bad for [NVIDIA] misses the load shape. A1.05T-parameter MoE that activates49B per pass is the wrong shape to run on commodity accelerators; it depends on very-high-bandwidth memory, NVLink-class interconnect, and a software stack (TensorRT-LLM, Dynamo) that closed-weight APIs depend on even more. [NVIDIA]'s Q2 FY2027 — revenue $96.2B and net income $59.7B on a 75% gross profit margin, reported on the 10-Q filed August 26, 2026 — shows the data-center demand curve already steepening independent of [Mistral]. Add [Mistral]'s 3,800 Grace Blackwell training cluster plus the Microsoft-funded Vera Rubin expansion, and [NVIDIA] collects the largest single new open-weight workload of the cycle, plus every inference hour enterprises pull rather than route through a closed-API discount.

Short term (days to quarters), the catalyst tape is dense: Mistral Large 4 weights land on Hugging Face by end of October; benchmark churn (DuetBench, AA Intelligence Index) will refresh through Q4; enterprise RFPs that priced closed APIs in September will re-bid. The immediate winners are [NVIDIA] (more silicon), [Microsoft] (Azure sovereign-AI differentiation) and [Amazon] (Bedrock gains an open-weight anchor). The immediate losers are [Alphabet]'s Gemini premium tier and [Meta]'s Llama team, both of whom now compete against a 1T Apache 2.0 model with hyperscaler distribution. Longer term (1–3 years), the pricing floor for frontier inference resets under $1/MTok output — Claude Opus and GPT-6 Astra pricing either comes down materially or revenue growth decouples from token volume growth at [OpenAI] and [Anthropic].