Mistral Studio Lets Enterprises Run AI Agents Inside Their Own VPC
Mistral Studio, announced November 5, 2026 alongside Mistral Large 4, lets a company build and run AI agents in the cloud, in its own VPC, on-premises, or a mix — a frontier-scale open model shipped with an enterprise agent platform that lets the buyer decide where the work executes. The details are in Mistral’s announcement.
The platform bundles managed agents, with skills, code sandboxes, subagents, hooks and limits on the tools and data each agent can touch. Workflows combine fixed rules, agent reasoning and human approvals, with resumable runs. OpenTelemetry traces record model calls, tool use and approvals, and evaluations test a prompt, model or tool change against the customer’s own datasets before it ships. Mistral names ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services as production users, and CMA CGM says new automated processes are production-ready in weeks rather than months. The savings Mistral cites — hundreds of hours a month for an incident-investigation agent, about 35 hours a week for a product operations workflow — are the company’s own claims.
The model underneath is priced for self-hosting. Mistral Large 4 is a 1-trillion-parameter open-weight model at $0.70 per million input tokens and $2.10 per million output tokens. It runs on vLLM or SGLang, and its modified MIT license allows commercial use except for businesses offering model-as-a-service.
The outcome to weigh is control: where company data sits, who approves an agent’s action, and whether a change can be tested before it reaches production. Studio puts all three inside the platform, and the open weights keep the exit open. Before wiring any agent to company systems, ask every vendor for the same three answers.