Mistral expands Studio to build and run AI agents

What's new

On November 5, 2026, Mistral announced an expansion of Studio, the company's platform for building AI solutions in businesses. The news is tied to the flagship model Mistral Large 4, known as Le Chonk, which is open weight. That means the model's "weights" are open and can be downloaded.

Studio rests on three building blocks. Agents are AI assistants that make decisions on their own, and they are built in Python with the Agents SDK. Workflows are fixed processes with state, retries, schedules and human steps. Apps are finished programs for users with built-in login and hosting. New additions include a workflow builder where you can describe the process in plain language in Mistral Vibe, a command line for apps, a managed search index, webhooks and triggers, and tools for monitoring and testing the agents.

What's smart

The platform bundles what companies otherwise have to stitch together themselves. The AI systems get access to company data through connectors, libraries and search indexes. Role-based access control and moderation rules apply across the entire platform.

Standard tracing, essentially a logbook of what an agent has done, automatically hides secrets and personal data. There are also evaluations, where you test agents' responses both along the way and in advance with managed datasets. By default, an agent session has a 24-hour lifetime (adjustable), closes after 4 hours of inactivity and waits at most 1 hour for an approval.

Agents and workflows can run in Mistral's cloud, at the customer's own site or locally during development. With self-hosting, the customer runs workers on its own infrastructure, while Mistral handles the orchestration. Apps, however, can only run in Mistral's cloud.

Cheaper or better?

Pricing for the new features has not been published, and the documentation mentions no price per run or hour. Managed workflow deployments require a pay-as-you-go or Enterprise plan, and the free plan cannot be used for them. Tracing and evaluations are only for Enterprise organizations.

According to Runtime Wire, Mistral is positioning Studio against OpenAI, AWS and Cohere, among others, with open weights and flexible deployment as selling points and regulated customers as the target audience. There are no independent measurements of performance or time savings. Nearly everything is also in public preview, so interfaces and behavior may change. Only Studio apps in Vibe Code Web are exempt, while HTTP API connectors are in private preview. The older Agents API and durable agents are marked for phase-out, with no announced end date.

The sources do not say whether or when the new features will come to private cloud and on-premises, or what availability looks like in Denmark and the EU.

What it's good for

Studio is aimed at larger and regulated companies that want to use AI agents in workflows where data, permissions, approvals and auditing must be under control. Runtime Wire points out that companies' AI spending is shifting from access to models toward the very systems that tie everything together.

According to Mistral, ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services, among others, already use Studio in production. For an ordinary user, the platform is less relevant. For a company with IT staff who want to build their own agents with human approval and traceability, it is worth keeping an eye on once pricing arrives.