French AI startup Mistral has unveiled its latest frontier-level language model, Mistral Medium 3, marking a significant advancement in enterprise AI. According to the Paris-based company, this new model outperforms leading systems like GPT-4o and Claude Sonnet 3.7 across various benchmarks, while offering significantly lower operational costs.
Medium 3 is designed to combine high performance with affordability and flexibility, making it suitable for a wide range of business applications. The model is accessible via Le Chat Enterprise, Mistral’s new AI assistant platform. This tool includes features such as an agent builder and seamless integration capabilities with enterprise software ecosystems. Mistral has also hinted at the upcoming release of an even more powerful model in the near future.
One of Medium 3’s key strengths lies in its cost-efficiency. The company claims the model is eight times cheaper to run compared to many of its rivals. Specifically, it operates at a rate of $0.40 per million tokens for input and $2 per million tokens for output, making it more affordable than alternatives like DeepSeek V3.
Tailored for enterprise deployment, Medium 3 supports hybrid and on-premises infrastructure, including self-hosted environments running on just four GPUs. It also allows custom post-training and can be integrated directly into enterprise workflows and tools.
Performance-wise, Medium 3 has shown exceptional results in core professional tasks. It leads in benchmarks like ArenaHard (97.1%) for instruction following and Math500 (91%) for mathematical reasoning. In long-context processing tests such as RULER 32K, it achieved an impressive 96%. According to Mistral, it consistently outperforms other top models like LLaMA 4 Maverick, Cohere Command R+, and DeepSeek 3.1, particularly in coding and human-evaluated tasks.
With Medium 3, Mistral positions itself as a serious contender in the global AI space, aiming to deliver frontier-grade capabilities with broader accessibility and enterprise readiness.
