Mistral Continues Rapid Iteration
Mistral AI, the Paris-based AI startup valued at €5.8 billion, released Mistral Large 2.5 this week — the fourth major model update in 18 months, reflecting the company's strategy of rapid iteration rather than large infrequent releases. The model shows measurable improvements in coding benchmarks and features a standard 128K token context window.
Benchmark Results
On HumanEval coding benchmarks, Mistral Large 2.5 scores 88.3%, up from 84.1% in Mistral Large 2. On MATH-500, the model scores 89.2% (up from 85.7%). Multilingual performance on European languages — a historically strong point for Mistral — improves further, with particularly strong results in French, German, Italian, Spanish, and Polish.
Pricing
Mistral Large 2.5 is available via Mistral's La Plateforme API at €6 per million input tokens and €18 per million output tokens — approximately 20% less than comparable Claude 3.5 Sonnet and GPT-4o pricing. The company continues to position on price-performance rather than absolute performance leadership.
Enterprise Focus
The release comes alongside expanded enterprise features including an EU data residency option with data processing entirely within French data centers, FedRAMP-equivalent certification targeting European public sector customers, and extended rate limits for enterprise contracts.
Strategic Context
Mistral occupies an important position as the leading European AI foundation model company. Its open-weights releases (Mistral 7B, Mixtral 8×7B) have been widely adopted by the open-source community, while its commercial API serves European enterprises that prefer not to route data through US cloud providers.