Mistral Vibe

Mistral Vibe is no longer the relatively simple Le Chat assistant that Mistral launched in 2024. In May 2026 Mistral folded Le Chat and its coding-agent line into one product family: Vibe for work and Vibe for code. The current product spans ordinary chat, web search, Deep Research, voice, images, documents, persistent work context, more than 100 connectors, long-running Work tasks and coding agents that can operate on the web, in an IDE or from the terminal.¹²

That makes Vibe much closer to ChatGPT Work or Claude/Cowork than to a conventional chatbot. The important product distinction is that Vibe uses Mistral's own model family and can also be deployed in ways the largest U.S. consumer assistants generally do not offer: enterprise customers can run it on Mistral Cloud, in a private cloud or on their own infrastructure, with regional/data-residency options and custom models.¹

The current default model is Mistral Medium 3.5, a dense 128B open-weight multimodal model with a 256k-token context window. Mistral says Medium 3.5 replaced earlier separate Le Chat reasoning models and Devstral 2 in Vibe, consolidating ordinary instruction following, reasoning, coding and agentic work into one model.³⁴

This is an AiToolMap desk review. We did not create an account, connect business systems, run our own Vibe coding session or claim hands-on testing. Product-level judgments below come from current first-party documentation, a current multi-week TechRadar Pro test, current app-store evidence and the fixed AiToolMap panel.

Vibe's Work mode is designed for multi-step tasks rather than isolated prompts. Mistral describes it as an agent that can catch up across inboxes and calendars, perform deep research over public and internal sources, analyze spreadsheets and connected resources, draft reports and coordinate recurring workflows.²¹

The connector layer is broad. Mistral advertises more than 100 integrations and full MCP compatibility, including email, calendars, Slack, GitHub, Jira and other enterprise tools.¹ Its Google connector documentation is unusually concrete: Vibe can search and read Gmail, create drafts, access Calendar, create/update events, search Drive and read files. Write operations require explicit user approval, and authentication remains with Google's OAuth flow rather than Vibe storing the user's password.¹⁵

This approval model is important. The useful part of an agent is that it can act; the dangerous part is exactly the same thing. Mistral's explicit per-action approval rule for write operations gives users a clearer trust boundary than a vague promise that the agent will “be careful.”

TechRadar's hands-on test supports the product claim without suggesting full autonomy. The reviewer ran research, document and connected-tool tasks over several weeks and found Work Mode capable of handling multi-step research and producing structured outputs in one run.⁵ That is stronger evidence than merely listing “agents” on a feature page.

Vibe for code is the other half of the product. Mistral supports terminal use, VS Code and JetBrains integrations, background cloud agents and a “teleport” workflow that moves a session between remote and local environments.¹⁴

Mistral Medium 3.5 is now the default model for these long-running code tasks. Mistral's model card describes a single 128B model intended to cover instruction following, reasoning and coding; the current model documentation lists structured outputs, function calling, agents/conversations and document Q&A.³

TechRadar tested a GitHub-connected coding workflow and reports that Vibe wrote, tested and prepared a draft pull request in its sandbox, though the reviewer intervened twice with additional direction.⁵ That is a useful reality check: the coding agent is capable of completing an end-to-end unit of work, but “agentic” does not mean “unsupervised software engineer.”

The Register adds narrower context around Leanstral, a proof-oriented coding agent available inside Vibe. The publication reports Mistral's claims of competitive formal-verification performance at much lower cost than some Claude configurations, while correctly presenting those figures as Mistral's own benchmark claims rather than independent validation.¹³

The most relevant independent panel result comes from Artificial Analysis. Its current Mistral Medium 3.5 comparison reports an Intelligence Index of 30, a 256k context window and output speed around 140 tokens/second.⁷

This is a respectable result, but it matters that the current frontier has moved higher. Medium 3.5 is not in the top cluster of the Artificial Analysis Intelligence Index. The product therefore cannot rely on “best model” positioning in the way a leading Claude or GPT configuration sometimes can.

LM Arena tells a similar story from a human-preference perspective. Its official August 19 dataset export places Medium 3.5 around rank 41 overall with roughly 6,018 votes.⁸ Again: credible, current and competitive, but not frontier-leading.

The model has strengths those single-number rankings do not fully capture. It is open-weight, multimodal, reasoning-capable and designed for tool use. The official Hugging Face model card confirms 128B parameters, 256k context and Mistral's published 77.6% SWE-bench Verified result.¹⁶ AiToolMap treats the 77.6% number as a vendor/model-card result because the panel's official SWE-bench search did not independently expose a current Medium 3.5 row during this cycle.

Mistral also continues to improve the tooling around the model. Its August Agentic Search release reports large gains from iterative retrieval and navigation on FinanceBench and OfficeQA Pro when Medium 3.5 is paired with Mistral's search toolkit.¹⁷ These are first-party results, useful for understanding the product architecture but not independent benchmark votes.

Mistral's current public pricing page lists:

• Free: $0, web/mobile access, limited messages/searches/coding sessions, image generation, 100+ connectors and $10/month API credit.
• Pro: $14.99/month before tax, more usage, complex long-running tasks, all-day coding, more images, $30/month API credit and chat/email support.
• Team: $24.99/user/month before tax, collaborative workspace, up to 30GB storage per user, domain verification and data export.
• Enterprise: custom pricing with private deployment, custom models/agents/workflows, audit logs, SAML SSO and white labeling.¹⁴

The Pro price is below the $20 reference tier of ChatGPT Plus, Claude Pro and Google AI Pro. That makes Vibe unusually easy to justify as a second assistant for users who specifically value Mistral's model/open/deployment philosophy.

The problem is quota transparency. Mistral's pricing page uses relative language—“limited,” “more,” “up to 5x”—rather than a stable public request budget. TechRadar independently flags the same issue.¹⁴⁵ Value therefore depends not just on the monthly price but on how quickly an individual's real Work/Code tasks hit fair-use limits.

API economics are separate. Mistral currently prices Medium 3.5 at $1.50 per million input tokens and $7.50 per million output tokens.³ This is competitive but not the cheapest model in the market; the value case is stronger when a user wants the same model to cover ordinary reasoning, multimodal work and agentic coding.

Google Play is the largest current external user sample AiToolMap could substantively read. The current Italian listing reports about 4.1/5 from roughly 43,900 reviews and more than one million downloads.⁹ That is a meaningful sample, but its aggregate is noticeably below the roughly 4.5–4.8 range seen on several leading mass-market AI assistants.

The visible 2026 reviews are mixed. Positive users describe Vibe as useful, inexpensive and capable for everyday work. Negative themes include instruction-following failures, latency/stuttering, restrictive limits, connector problems, support delays and interface changes.⁹

The App Store evidence is much smaller in the directly readable storefront captured here: Ireland shows 4.8/5 from only 71 ratings.¹⁰ AiToolMap retains it as a mobile-surface signal but will not let such a small sample compete numerically with the Google Play population.

G2 is even smaller. Its current Mistral/Le Chat profile has one verified-current-user review and a 5/5 aggregate.¹¹ One review is evidence that a real review surface exists, not evidence of population-level satisfaction.

Trustpilot points in the opposite direction: 2.5/5 from 96 reviews, with 49% one-star and 62 reviews posted in the prior 12 months.¹² Recent complaints mention answer quality, limits, outages, connectors, billing and support; positive users praise coding and the European alternative. Under AiToolMap methodology, Trustpilot is qualitative-only because self-selected customer-service populations should not become a direct “product quality” score.

The important conclusion is that Vibe's user evidence is currently more polarized and much smaller than the evidence bases for ChatGPT, Claude or Gemini.

Mistral's current privacy documentation says Free Vibe conversations may be used to improve models unless the user opts out. Pro, Team and Enterprise conversations are not used for model training by default. API data is not used for model training. Uploaded Library documents are not used for training.¹⁸

Connector data receives additional protections. Mistral says normal connector data is fetched on demand and is not permanently stored, and connector content is not used for training.¹⁸ There is an important nuance: the older Google Drive/SharePoint “Knowledge Connector” design does index and store files in European data centers for search. Mistral is deprecating that architecture at the end of August 2026 in favor of MCP connectors.¹⁹

That transition is exactly the kind of detail a privacy-sensitive buyer should care about. “Connectors don't store data” can be true for the new architecture while an older indexing connector still has a stored index during migration.

Vibe's optional Memories feature can also store user preferences across conversations, but users can view, add, edit, delete or disable those memories.²⁰

For enterprise customers, Vibe can be deployed on premises, in a private cloud or on Mistral Cloud with data-residency controls.¹ This is a structural advantage for organizations that cannot simply send all sensitive workloads to a U.S.-hosted consumer service.

Vibe's design includes explicit approval for write actions on connectors such as Google Workspace.¹⁵ That is a sensible safeguard and should remain central to the product as agents become more autonomous.

Still, independent user evidence shows that ordinary reliability remains uneven. Google Play reviews include complaints about instruction following and latency; Trustpilot includes more severe complaints around repeated mistakes and service issues.⁹¹² These are not controlled evaluations, but they matter because an agentic product compounds ordinary model errors with tools and permissions.

Mistral's ability to run Vibe locally or on private infrastructure can reduce some data-exposure risks, but it does not remove model-output risk. The company's own offline-model documentation warns users to validate outputs on representative tasks and to understand which external MCP tools or connectors can still send data outside the machine.²¹

Vibe is most compelling for users who want a capable general assistant plus coding/agentic work but do not want to default automatically to a U.S. hyperscaler ecosystem.

It is particularly worth testing for European professionals, developers and organizations that care about regional deployment, open weights, private infrastructure or model-training defaults. The $14.99 Pro tier lowers the cost of experimentation.

It is a weaker fit for users who want the strongest currently benchmarked model regardless of price, teams deeply embedded in Microsoft 365 or Google Workspace, or buyers who need a large body of independent customer evidence before procurement.

Free is useful for initial orientation but is intentionally constrained. Pro is the relevant individual tier. Team should be evaluated mainly for organization/storage/admin needs rather than assumed to deliver a proportionate jump in model quality. Enterprise is where Mistral's sovereignty and deployment advantages become most differentiated.