By ByteBot
Mistral shipped a full AI coding agent in May 2026. Not a renamed chatbot — a proper multi-surface agent with a VS Code extension, terminal CLI, web Code Mode, and cloud Remote Agents. It runs on Devstral 2, an open-weight model hitting 72.2% on SWE-bench Verified, and Mistral Medium 3.5, a 128B model priced at roughly half what you pay for Claude Sonnet. If you’ve been waiting for a credible open-weight alternative to Claude Code, Vibe is the most serious contender yet.
What Changed: Le Chat Became Vibe
Mistral’s consumer chatbot Le Chat is gone. In its place is Vibe — one product, one subscription, covering both productivity work and software development. The rebrand landed May 22, 2026, alongside capabilities that turned a chat interface into a full agent platform. Every existing Le Chat account, setting, and conversation carried over automatically.
Vibe operates across four distinct surfaces:
- VS Code extension — A side-panel agent that reads, edits, and runs commands across your entire project. Open files attach automatically; @ mentions pull in context from other directories. GitHub, GitLab, Jira, and Linear integration is built in.
- Terminal CLI — An open-source Python package installable in seconds. Bring your own key or use a Mistral plan. Fully configurable via config.toml.
- Web Code Mode — A browser-based coding environment connected to GitHub. Sessions run in isolated sandboxes, ship reviewable PRs, and persist while your machine is off.
- Remote Agents — Cloud-executed sessions that run independently, notify you on completion, and open pull requests. Slack triggers are coming.
The /teleport command moves a live terminal session to the cloud and back, preserving history and approvals intact. No other major coding agent does this today.
The Open-Weight Case
This is Vibe’s strongest argument. Devstral 2, the 123B model powering the agent, is the highest-scoring open-weight coding model on SWE-bench Verified at 72.2%. It ships under a modified MIT license — you can run it yourself, with no API calls and no code leaving your infrastructure.
Devstral Small 2, the 24B variant, scores 68.0% on the same benchmark and runs on consumer hardware under Apache 2.0. For a team with capable GPUs and a strict data policy, this is a serious option — fine-tune on your own codebase, self-host the inference, pay nothing per token.
For regulated industries, EU-based companies, and enterprises that refuse to send proprietary code to third-party APIs, the open-weight story here is the most compelling in the market. Cursor and Claude Code offer nothing comparable.
Getting Started With the CLI
Install takes under a minute:
pip install mistral-vibe
# Or with uv (faster installs)
uv tool install mistral-vibe
# Run in your project directory
cd your-project
vibeThe setup wizard runs on first launch. Sign in with a Mistral account or provide an API key. Configuration lives at ~/.config/mistral-vibe/config.toml — model, provider, tool permissions, and UI preferences are all adjustable there. Skills (slash commands for repeatable workflows) and custom subagents can be defined in the same file.
Pricing: Free tier included. Pro is $14.99/month. Team plans run $24.99/user/month with shared workspaces and admin controls.
The Cost Argument
Mistral Medium 3.5 is priced at $1.50 per million input tokens and $7.50 per million output tokens via API. Claude Sonnet runs $3/M input and $15/M output. The math is straightforward: roughly 50% lower cost at scale.
For high-volume agentic workloads processing thousands of files or running continuous CI loops, that gap compounds fast. Worth modeling before you dismiss Vibe as “not as good as Claude.”
Where It Falls Short
Vibe is a real product. It is also not beating Claude Code right now, and saying otherwise would be dishonest.
Human evaluation still favors Claude Sonnet 4.5 over Devstral 2 — Vibe’s 42.8% win rate is meaningful but not dominant. The context window is 256K versus Claude Fable 5’s 1M token ceiling, which matters for large codebases analyzed in a single pass. The VS Code extension lacks Cursor’s tight native-IDE feel and the ability to hot-swap models mid-session.
The EU sovereignty angle also deserves a caveat: independent audits have noted that some Vibe data processing still routes through US infrastructure. “EU-hosted” is closer to “EU-preferred” at this point. Teams with strict data residency requirements should verify specifics with Mistral directly before committing.
Should You Switch?
The honest answer depends on your situation:
- Switch (or add): You have EU compliance requirements, want self-hosted inference, run high-volume agent workloads where cost matters, or need async PR-shipping workflows without babysitting a terminal session.
- Stay put: You rely on 1M context windows, need maximum model performance for complex reasoning tasks, or want the most polished single-vendor experience. Claude Code or Cursor still wins there.
- Run both: Practical answer for many teams. Use Vibe Remote Agents for async ticket work, Claude Code for interactive complex sessions. The two aren’t mutually exclusive.
Mistral Vibe isn’t the default choice for most developers yet. But it’s now in the conversation in a way that Le Chat never was. The open weights, the cost curve, and the remote execution model give it a real claim on a specific workload category — and that category is growing. Check the independent coding agent benchmarks before making a final call.
