Mistral's new Studio managed agents, workflows and apps are a public preview hosted by Mistral, with no published price. If you are choosing Mistral over AWS, Microsoft or Google because your data must stay on infrastructure you control, the November 5 release does not yet give you that by default. Self-hosted agents still hand session state, routing and traces to Mistral, apps run only on Mistral Cloud, and the fully on-premises Studio is a separate Enterprise contract.
Pilot it on data you could already send to a SaaS vendor. Hold production for regulated workloads until Mistral puts a price, a GA date and a self-hosted control plane in writing.
What Mistral Shipped in Studio on November 5
Mistral added a managed agent runtime, a workflow builder, hosted apps, tracing and evaluations to Studio, its enterprise AI platform. The announcement describes agents with skills (written team procedures), code sandboxes, subagents for delegating tasks and hooks for custom logic, with Studio running the runtime. Workflows mix fixed rules, agent reasoning and human approvals, can resume after an interruption and can start from events in enterprise systems. Apps get hosting, authentication and access controls and deploy with one command. Runs emit OpenTelemetry traces covering model calls, tool use and approvals, and you can test a prompt, model or tool change against your own datasets before release.
The announcement leads with Mistral Large 4, though customers choose their own models. We covered Large 4 when Mistral released its open weights. Builders can generate a workflow from a plain-language description inside Mistral Vibe, and a new Mistral MCP server lets other coding agents build and deploy to Studio.
Mistral names seven production users: ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and TCS. Two of them are European banks, and their quotes in the post are the reason sovereignty buyers will read this launch closely. BNP Paribas says "we run Mistral Studio on our own infrastructure and keep those systems under our control." ABANCA calls Studio "our platform for building enterprise AI applications," and an Italian trade write-up reports it runs Studio in its own cloud and uses only Mistral's inference.
Which Parts Are Preview, and Which Are GA?
Almost everything is public preview, and every new piece is labelled Mistral-hosted. Mistral's own release notes list 12 entries dated November 5, 2026:
| Feature | Status | Hosting |
|---|---|---|
| Managed Agents (SDK) | Public preview | Mistral-hosted |
| Workflow builder | Public preview | Mistral-hosted |
| HTTP API connectors | Private preview | Mistral-hosted |
| Mistral Apps CLI | Public preview | Mistral-hosted |
| Offline evaluations, managed datasets, datasets from telemetry | Public preview | Mistral-hosted |
| Managed search index | Public preview | Mistral-hosted |
| Tracing for managed agents | Public preview | Mistral-hosted |
| Webhooks and triggers | Public preview | Mistral-hosted |
| Online evaluations | Public preview | Mistral-hosted |
| Studio apps in Vibe Code Web | GA | Mistral-hosted |
Public preview is a real constraint. The Managed Agents docs warn that the "API, behavior, and limits can change during the preview," which means code you write now may need changes before GA. The one connector that would save integration work, registering any web API without first building an MCP server, is private preview and invitation only.
Teams already on Mistral's older agent API should check their code. DataStudios reports that the Agents and Conversations API (client.beta.agents) now sits under a Deprecated section of the docs, with no removal date.
Can You Run Mistral Studio Agents On-Premises?
You can run the agent harness and your tools on your own servers today, but Mistral still keeps the session. The Managed Agents docs describe three run modes: in-process for development, agents deploy where the harness and tools run on Mistral, and agents serve, where "the harness and your tools run on your servers" while Mistral stores the session so it survives a crash. A session records every message, tool call and result. For a bank, that record is the sensitive data.
The self-hosted deployments page spells out what stays with Mistral when you host the workload yourself: routing work to the right deployment, durability across crashes and restarts, load-balanced scaling, and the executions, logs and traces you see in Studio. It also says plainly: "Self-hosting a deployment is not the same as running the whole Studio platform on your own premises." Agents and workflows can be self-hosted this way. Apps cannot: they "run only on Mistral Cloud."
Workflows follow the same hybrid design. Your workflow code runs on workers you control, and those workers connect outbound only, while Mistral runs the orchestrator that holds state, history and task dispatch. Two details soften this. The SDK encrypts payloads before they leave your worker, so Mistral stores ciphertext, and inputs or outputs over 2MB go to your own S3, GCS or Azure blob storage with only a reference kept. Durable execution runs on Temporal, the engine we compared with its rivals in our durable execution cost piece.
Tracing is on by default. The release notes say agent, model and tool spans go to Studio automatically, with redaction applied before export. Check what that redaction removes before you point an agent at customer records.
The full on-premises option exists, but it is a different product. The workflows docs say Enterprise customers can run the Studio environment in their private cloud or on-premises, and Mistral's pricing page says custom Enterprise deployments can be self-hosted, private cloud or on-premises, priced on a "Contact us" basis. Mistral does not say which of these arrangements BNP Paribas uses. Neither the announcement nor the release notes say whether the November 5 features are in the self-hosted Studio build yet, or when they will be. The trade coverage disagrees on this: the Italian write-up says the release made everything installable on customer servers, while DataStudios read the same release notes as giving no self-hosted date. The docs support DataStudios.
What Does Mistral Studio Cost?
Mistral has published no price for managed agents, workflows, apps or evaluations. None of the November 5 docs give a per-run, per-hour or per-trace figure, and the pricing page (checked November 7, 2026) lists plans, not runtime meters. Pro is $14.99 a month with $15 of API credits, Team is $24.99 per user per month, and Enterprise has no listed price. The only model rate shown is Mistral Large at $0.50 per million input tokens and $1.50 per million output tokens.
Plan gates matter more than the sticker. The Apps docs say Free users, and Pro or Team users without pay-as-you-go billing, can see the deploy commands but cannot create apps. DataStudios reports that managed workflow deployments need pay-as-you-go or Enterprise, and that traces and evaluations are listed for Enterprise organizations.
Compare that with the runtimes Mistral is competing against, which do publish meters. Anthropic prices Claude Managed Agents at $0.08 per session-hour, counted only while the session is running, plus tokens, and charges 1.1x on tokens when inference is pinned to the US. AWS lists AgentCore Runtime at $0.0895 per vCPU-hour and $0.00945 per GB-hour on its v1 meter, with memory, gateway and observability billed separately. Our runtime pricing comparison worked those meters into a monthly bill for Google, AWS and Microsoft. You cannot do that arithmetic for Mistral yet, which makes a budget request for a production rollout guesswork.
Why Sovereignty Buyers Should Still Take It Seriously
Mistral sells the whole agent platform for installation on hardware you own, which the hyperscalers do not. The Italian write-up's comparison places AWS's AgentCore, Microsoft's Copilot Studio and Google's agent platform inside each vendor's own cloud, and notes that Anthropic's managed agents can run sandboxes on your servers while the loop that decides the agent's steps stays on Anthropic's. Mistral's Enterprise route puts the Studio environment itself in your private cloud or data centre, and BNP Paribas says it runs Studio on its own infrastructure.
Encrypted payloads, outbound-only workers and your own blob storage also give the hosted hybrid mode a better data story than most SaaS agent platforms. For workloads where metadata and orchestration state can leave the building, that may already be enough.
The gap is narrower and more specific: the newest features are documented only as Mistral-hosted, with no published date for the self-hosted build and no price. This is the same pattern we flagged in August, when Mistral's European Compute Units offering left agent features off the regional endpoint. Score the contract terms, as we argued after the Series D, and treat the flag on the company as secondary.
What to Do Before You Sign
This Week:
- Ask your Mistral account team, in writing, which of the 12 November 5 features are in the self-hosted Enterprise Studio build today, and the date for any that are not.
- Have your data protection lead read the session and tracing behaviour for
agents serveand decide whether a Mistral-held session record is acceptable for each workload you have in mind. - If you call
client.beta.agents, find every call site and open a migration ticket now, while there is no removal date.
This Month:
- Run a pilot on Mistral-hosted Studio with data already cleared for SaaS processing. Turn on the OpenTelemetry export and confirm what the default redaction actually strips from a trace.
- Build one workflow on self-hosted workers with payload encryption and your own blob storage, and check what still appears in Studio's history.
- Price the same pilot on Amazon Bedrock AgentCore or Claude Managed Agents using their published meters, so you have a number to hold Mistral's quote against.
Before Renewal:
- Get the runtime meter, the trace retention period and a GA date into the order form, and add a clause that preview features moving to GA cannot change price without notice.
- If on-premises is your reason to buy, make delivery of the self-hosted build a condition of the production commitment.
The Bottom Line
Studio now has the pieces buyers expect from an agent runtime (sandboxes, durable workflows, hosted apps, traces, evals), and two European banks are on the record using it. Mistral has not yet published the part regulated buyers are paying for: a price, a GA date and documentation showing the new features running entirely inside your perimeter. No Mistral document we found gives a date for self-hosted versions of the new features, so plan as if there is none until your account team writes one down. Until then, run Studio's managed agents as a pilot.
Continue Reading
- Google Agent Runtime's $510 Meter Beats AgentCore's $846
- LLM Data Residency: Which Providers Actually Keep Data In Region
- Mistral Wants Multi-Year Money. Its EU Endpoint Drops Agents.
- Claude Managed Agents: Why Anthropic Runs Your AI Infrastructure
- Microsoft Starts Billing for Copilot Harness on December 1
- Why Temporal Lost to Mistral in the $2B AI Orchestration War