Mistral Large 4 Launches: Pricing, Benchmarks and Open Weights

Mistral AI released Mistral Large 4 on October 6, positioning the new 1-trillion-parameter system as a cost-efficient alternative to leading proprietary models while maintaining an open-weight licensing strategy. The Paris-based company priced the model at $1.36 per million input tokens and $4.18 per million output tokens, significantly undercutting competitors like Claude Opus 5.5 and GPT-6 Astra. This pricing structure targets enterprise users seeking high-capacity reasoning without the premium costs associated with current frontier closed-source APIs.

Despite its total parameter count, Large 4 operates using a Mixture of Experts (MoE) architecture that activates only 49 billion parameters per query. This sparse activation mechanism allows the model to deliver performance comparable to much larger dense models while keeping inference latency and computational overhead manageable. The previous flagship, Large 3, utilized a similar approach with 41 billion active parameters out of 675 billion total. By increasing the active parameter count by roughly 20%, Mistral aims to close the capability gap with top-tier U.S. and Chinese systems while preserving the economic advantages of efficient routing.

Benchmark results indicate Large 4 occupies a competitive middle ground in the current AI hierarchy. On the Finance Agent v2 evaluation, it scored 54.7, surpassing GPT-6 Astra’s 53.5 but trailing Claude Opus 5.5’s 58.6. In blind human coding evaluations conducted by Surge AI, Large 4 achieved a 3.74 out of 5, placing it second among tested models but behind Claude Opus 5’s 4.22. For broader business automation tasks measured by AutomationBench, Large 4 scored 59.9, which lags behind Gemini 4 Argon (77.5) and Claude Sonnet 5.5 (71.8). These metrics suggest the model excels in specialized financial and coding contexts but faces stiffer competition in generalized agentic workflows.

The release serves as the first major technical milestone following Mistral’s €3 billion ($3.37 billion) Series D funding round in September, which valued the company above €21 billion. Samsung led the investment, signaling strong institutional confidence in Mistral’s roadmap. Unlike many U.S. labs that have shifted toward closed ecosystems, Mistral continues to bet on open weights as a differentiator for “sovereign AI” deployments. This strategy appeals to governments and corporations requiring data residency and auditability, evidenced by a recent multi-hundred-million-euro agreement with Saudi Arabia’s state-backed HUMAIN entity.

Mistral has committed to releasing the full model weights by the end of October. This pending open release will allow independent researchers and developers to validate benchmark claims, fine-tune the system for niche applications, and deploy it on private infrastructure. Until then, the model is available via API, offering immediate access for integration testing. The community has already embraced the release under the nickname “Le Chonk,” a playful reference to an earlier internet meme that the company has since adopted officially, reflecting a distinct cultural branding strategy that differentiates it from more austere Silicon Valley counterparts.