Tools · · 2 min read

Mistral ships Vibe remote agents on Mistral Medium 3.5 — cloud coding agents at $1.5/$7.5 per 1M tokens

At its first AI Now Summit (Paris, May 28, 2026), Mistral launched Vibe — cloud coding agents that run async and open PRs — powered by Mistral Medium 3.5, a 128B dense model with a 256k context that scores 77.6% on SWE-Bench Verified at one-tenth the price of frontier US models.

Mistral used its first AI Now Summit (Paris, May 28, 2026) to ship Vibe — a unified agent platform that takes coding work “from request to merged change, across the web app, your editor, and your terminal” (Source: Mistral AI Now Summit). The headline builder feature is Vibe remote agents: cloud agents that run asynchronously, in parallel, and open reviewable pull requests on their own. They are powered by a new model, Mistral Medium 3.5.

Key facts:

  • Mistral Medium 3.5 is a 128B dense model with a 256k-token context window, in public preview under a modified MIT license. (Source: Mistral AI)
  • API pricing is $1.5 per 1M input tokens and $7.5 per 1M output tokens.
  • It scores 77.6% on SWE-Bench Verified and 91.4 on the τ³-Telecom agentic benchmark, per Mistral’s own numbers.
  • Remote agents integrate with GitHub, Linear, Jira, Sentry, Slack, and Teams, and can “teleport” a local CLI session to the cloud with history intact.
  • Vibe is available on Pro, Team, and Enterprise plans.
Mistral's official chart of Mistral Medium 3.5 agentic benchmarks (SWE-Bench Verified, τ³ Telecom/Airline/Retail/Banking, BrowseComp) versus Kimi K2.5, GLM 5.1, and Qwen3.5
Mistral Medium 3.5 (128B) agentic scores vs competing open-weight models: 77.6% SWE-Bench Verified, 91.4 τ³-Telecom, 48.6 BrowseComp. Footnotes matter — Mistral notes Sonnet 4.6 hit reasoning-truncation issues from API restrictions, GPT-5.2 ran at low reasoning effort, and τ³ uses a user simulator. (Source: Mistral AI)

What this means if you’re building with coding agents

Vibe remote agents drop into the same slot as Cursor’s background agents, Claude Code, and OpenAI’s Codex — fire off a task, walk away, review a PR later. The differentiator is price-per-capability. At $1.5 / $7.5 per 1M tokens, Mistral Medium 3.5 is roughly an order of magnitude cheaper than frontier US models on output tokens, while landing a 77.6% SWE-Bench Verified score that sits in competitive territory rather than at the top.

The honest read: the headline benchmarks are self-reported, and the footnotes are doing work. Mistral applies its own context-management and “discard-all at 100k tokens” strategy on the agentic suites, and several competitor numbers ran under constrained settings. Treat 77.6% as “competitive, verify on your repo” — not “beats Opus.” For a frame on how a cheaper cloud tier changes agent economics, see our Cursor Composer 2.5 guide and the Reasonix cache-engineered coding agent writeup.

Mistral chart comparing Mistral Medium 3.5 against previous Mistral models on agentic benchmarks
Versus Mistral's own prior models, Medium 3.5 jumps from 64.3 to 76.1 on τ³-Retail and 41.5 to 72.0 on τ³-Airline — a real generational step. (Source: Mistral AI)

The strategic context: multiple outlets (VentureBeat, Futurum, meshedsociety) report Vibe is a rebrand of Le Chat into a single agent spanning productivity and software development, alongside a full-stack “Mistral for Industrial Engineering” push with Airbus, BMW Group, and ASML, and a new 10 MW inference data center near Paris. For EU teams with data-residency constraints, a frontier-adjacent coding agent hosted in France — at one-tenth the output price — is a genuinely new option. The catch worth watching: no published per-task cost ceiling on remote agents, so a runaway agent loop is on your bill.

What’s still unclear

  • VS Code / IDE specifics: Mistral describes “Vibe for Code” across terminal, IDE, and background, but hasn’t published the exact extension/setup details.
  • Independent benchmarks: all coding scores are self-reported; no third-party SWE-Bench Verified run yet.
  • Rate limits on remote agents: parallelism is advertised but concurrency caps per plan aren’t documented.

Sources

Source: Mistral AI