French AI company Mistral AI is rolling out Mistral Medium 3, a new language model aimed at enterprise customers, alongside its business-focused assistant platform Le Chat Enterprise.
Mistral positions Medium 3 as a "medium-sized" model that, according to the company, achieves over 90 percent of the performance of Anthropic's Claude Sonnet 3.7, but at a much lower price point. Input costs run $0.40 per million tokens, with output at $2 per million tokens. That makes Medium 3 cheaper than previous price leaders like Deepseek v3, both for API access and self-hosted deployments. The ongoing drop in AI model pricing continues.
Mistral highlights Medium 3's strong performance in programming tasks and STEM benchmarks. Internal evaluations show the model outperforming open-source systems such as Llama 4 Maverick and even commercial offerings like Cohere Command R. Mistral says these tests relied on a mix of external benchmarks and in-house evaluations.
Medium 3 is designed for hybrid deployment, supporting on-premises or virtual private cloud setups and running on as few as four GPUs. Companies can further train the model, integrate it with existing infrastructure, and tune it for industry-specific needs.
The model is available now through Mistral La Plateforme and Amazon Sagemaker, with support for IBM WatsonX, NVIDIA NIM, Azure AI Foundry, and Google Cloud Vertex coming soon. Le Chat Enterprise is already live on the Google Cloud Marketplace, with more cloud integrations planned.
With Medium 3 now out and already ahead of many open-source flagship models, Mistral is hinting at an even larger model in development—suggesting there's still room for improvement. It remains Europe's only major AI model developer competing with US and Chinese companies in the large language model space.
Le Chat Enterprise brings AI tools to the workplace
Alongside its new model, Mistral is introducing Le Chat Enterprise, an AI assistant platform built on Medium 3. The platform is meant to address typical issues companies face with AI: scattered tools, insecure data connections, and "rigid models." Like enterprise offerings from OpenAI and Anthropic, Le Chat Enterprise currently only supports Mistral's own models.
Le Chat Enterprise offers company-wide search across sources like Google Drive, Sharepoint, and Gmail, automated agent creation with no coding required, and dedicated document libraries. Auto-summarization is built in to speed up content processing. Mistral says more connectors and an MCP interface are on the way.
The platform supports both cloud-based and self-hosted deployments. According to Mistral, access to internal company data is compliant with data protection standards, offering granular access controls and full audit logging. Companies can run the system within their own security domain and customize it as needed, including adding feedback loops or saving user reminders.
