Mistral AI launches Medium 3: High performance at eight times lower cost

Mistral Medium 3 achieves performance equal to or above 90% of Claude Sonnet 3.7 in general evaluations, but at a significantly lower cost of $0.4 per million input tokens and $2 per million output tokens. This model outperforms leading open models like Llama 4 Maverick and enterprise models like Cohere Command A, while in pricing it beats cost leaders like DeepSeek v3.

One of Medium 3's key advantages is its deployment capability on any cloud, including self-hosted environments with four GPUs or more. This flexibility addresses the needs of organizations that traditionally had to choose between fine-tuning via API or complete self-deployment from scratch.

In academic evaluations, Medium 3 particularly excels in coding and STEM tasks, approaching the performance of much larger and slower competitors. Third-party human evaluations confirm its superiority in the programming domain, offering better overall performance than some considerably larger competitors.

The model stands out for its adaptability to enterprise contexts. With Mistral's applied AI solutions, Medium 3 can be continuously pre-trained, fully fine-tuned, and integrated into enterprise knowledge bases. This makes it a high-fidelity solution for domain-specific training, continuous learning, and adaptive workflows.

Beta customers in sectors such as financial services, energy, and healthcare are already using the model to enrich customer service with deep context, personalize business processes, and analyze complex datasets. The Mistral Medium 3 API is available starting today on Mistral La Plateforme and Amazon Sagemaker, and will soon be available on IBM WatsonX, NVIDIA NIM, Azure AI Foundry, and Google Cloud Vertex.