Mistral's new model takes on Claude at a fraction of the cost

If you are tired of watching your frontier AI API bills climb, French AI lab Mistral just made a very direct bid for your enterprise budget.

The startup released Mistral Medium 3, a new model engineered to deliver near-frontier performance at a fraction of the operating cost.

Squeezing flagship performance into four GPUs

The headline here is the price-to-performance ratio. Mistral is offering Medium 3 via its API at $0.40 per million input tokens and $2 per million output tokens.

According to Mistral, the model performs at or above 90% of Anthropic's much pricier Claude Sonnet 3.7 across broad benchmark suites. It also beats recent open heavyweights like Meta's Llama 4 Maverick and Cohere's Command A.

Deployment flexibility is the other major hook. Mistral claims Medium 3 can run in self-hosted setups with just four GPUs or more, deploying easily across any cloud.

The company did not mince words on competition either, claiming Medium 3 beats cost leaders like DeepSeek v3 in both API and on-prem deployments.

Mistral says the model is tailored for coding, STEM problem-solving, and multimodal analysis. Beta testers across finance, energy, and healthcare—including BNP Paribas and AXA—have been putting it to work on customer service and internal workflow automation.

It rolled out Wednesday on Amazon SageMaker, with availability on Microsoft Azure AI Foundry and Google Cloud Vertex AI coming shortly. The lab also teased an even larger frontier model coming in the next few weeks.

Taking on enterprise workflows directly

Mistral also announced the general availability of Le Chat Enterprise, its direct answer to enterprise chatbot platforms.

Graduating from a private preview launched earlier this year, Le Chat Enterprise lets companies build custom AI agents and integrates directly into everyday tools like Gmail, Google Drive, and Microsoft SharePoint.

Notably, Mistral confirmed that Le Chat Enterprise will soon support Anthropic's Model Context Protocol (MCP). The open standard for connecting AI assistants to data repositories has gained serious momentum after endorsements from OpenAI and Google earlier this year.

Why it matters

For engineering leaders and teams spending heavily on frontier APIs, Mistral Medium 3 looks like an immediate candidate for pilot testing. If you run workloads that require private data handling or on-prem execution, getting near-Claude reasoning on a four-GPU footprint is a compelling proposition without relying solely on DeepSeek.

For the AI ecosystem, this release sharpens the squeeze on closed-source pricing power. Mistral is explicitly triangulating between Anthropic's elite intelligence and the rock-bottom pricing playbook popularized by open Chinese models.

We think Mistral's commercial playbook here is clever. Instead of getting bogged down in an unwinnable brute-force benchmark race, they are prioritizing deployment ergonomics, enterprise software integrations, and adopting open ecosystem standards like MCP. The lingering question is whether performing at 90% of a top-tier model holds up when real-world edge cases hit complex corporate workflows.