Mistral AI
European AI lab: open and proprietary LLMs, the Vibe assistant, and the La Plateforme API
Last reviewed 2026-06-20
Mistral AI is a Paris-based AI lab that builds large language models and the products around them. Its model family spans open-weight models released under permissive licenses (primarily Apache 2.0) and proprietary commercial models (Mistral Large, Medium, Small, plus coding models Codestral and Devstral and the Voxtral speech models), served to developers through its La Plateforme API. On top of those models it ships Vibe (formerly Le Chat), a consumer and enterprise AI assistant on web and mobile, and Studio (formerly AI Studio) for building and running agents and apps. Most of Mistral's user-facing surface is assistant- and copilot-grade: Vibe chats, searches the web, analyzes documents and images, generates images, and writes code on request. A newer, narrower slice is genuinely agentic. Vibe's Work Mode plans and runs multi-step tasks across connected tools, and Code Mode launches remote coding agents that open pull requests; according to Mistral, the agent maps out a plan and gets the user's sign-off before it starts and surfaces diffs for review, so the representative experience is a supervised agent under human direction rather than an autonomous one.
What it can do
Chat, search, and analyze with Vibe
AssistantVibe (formerly Le Chat) answers questions, searches the web, analyzes uploaded documents and images, and generates images on request across web and mobile apps.
sourceRun multi-step tasks in Work Mode
SupervisedWork Mode is an agentic mode that maps out a plan, gets the user's sign-off, then works through complex multi-stage tasks (catching up on communications, extracting data, drafting documents, running recurring workflows) using several connected tools.
sourceLaunch remote coding agents in Code Mode
SupervisedCode Mode launches remote coding agents (also available in a CLI, IDE, and VS Code extension) that write code and open pull requests; the user manages sensitive actions and inspects diffs as it writes.
sourceServe open and proprietary LLMs via La Plateforme
AssistantLa Plateforme is Mistral's developer API exposing open-weight models (Apache 2.0) and proprietary models (Mistral Large/Medium/Small, Codestral, Devstral, Voxtral) for pay-per-token use, with tool calling, OCR, code execution, and web-search add-ons.
sourceBuild and run agents in Studio
SupervisedStudio (formerly AI Studio) lets developers build, test, and run AI agents and apps on Mistral models; the lab also offers Forge for custom-model training and Compute for infrastructure.
source
Strengths
- +Open-weight models under permissive Apache 2.0 licenses plus proprietary commercial models in one lab
- +Vibe assistant, Studio agent builder, and the La Plateforme API span consumer, developer, and enterprise needs
- +Work Mode and Code Mode add real agentic work with explicit plan sign-off and diff review; self-hosting is available
Limitations
- −Most of the Vibe assistant surface is assistant/copilot grade; the agentic modes are newer and narrower
- −Recent Le Chat to Vibe rebrand (May 2026) means naming and feature docs are still settling
- −Frontier benchmark leadership is contested by larger US labs
Overview
Mistral AI is a Paris-based AI lab founded in 2023 by Arthur Mensch, Guillaume Lample, and Timothee Lacroix. It builds large language models, both open-weight (released under permissive licenses, primarily Apache 2.0) and proprietary, and the products around them: the Vibe assistant (formerly Le Chat), the La Plateforme developer API, and Studio for building agents and apps. It has raised substantial European funding, including a EUR 1.7B Series C in September 2025 led by ASML.
What it does
For end users, Vibe is the assistant surface: it chats, searches the web, analyzes uploaded documents and images, generates images, and writes code across web and mobile. A newer agentic slice goes further. Work Mode plans and runs multi-step tasks (catching up on communications, extracting data, drafting documents, running recurring workflows) across connected tools, and according to Mistral the agent maps out a plan and gets the user's sign-off before it starts. Code Mode launches remote coding agents (also in a CLI, IDE, and VS Code extension) that write code and open pull requests, with diff review and control over sensitive actions. For developers, La Plateforme serves Mistral's open and proprietary models pay-per-token, and Studio lets teams build and run agents and apps on top.
Integrations & setup
Vibe's Work Mode connects to Google Workspace, Outlook, SharePoint, Slack, and GitHub, and supports custom connectors and skills built on open standards. The La Plateforme API supports tool/function calling plus OCR, code execution, and web-search services. Several models are open-weight and can be self-hosted or deployed on-prem; enterprise deployment and white-label options are offered.
Pricing
Freemium. Vibe has a free tier (limited messages, web searches, and coding sessions), Pro at $14.99/month (expanded allowances and all-day coding across CLI, IDE, and web), and Team at $24.99 per user/month (collaborative workspace, storage, domain verification, data export). Enterprise is custom (custom models, agents, audit logs, SAML SSO, white-label). The La Plateforme API is pay-per-token with a free experimentation tier; add-ons such as OCR (around $2 per 1,000 pages), code execution, and web search are billed per call.
Best for / not for
Best for developers who want open-weight and proprietary LLMs from one European lab with self-hosting options, and for individuals and teams who want a capable assistant with optional agentic Work Mode and Code Mode. Less suited to buyers who need a mature, fully autonomous end-to-end agent today, or who want the single most benchmark-leading frontier model.
Alternatives
Claude, ChatGPT, and Google Gemini are the direct general-assistant and LLM-platform rivals; Ollama overlaps for running open-weight models locally. Mistral differentiates on its mix of permissively licensed open models, European data residency and self-hosting, and a unified Vibe agent across work and code.
What people are saying
We aggregate real LinkedIn discussion into sentiment for the agents people search most. Mistral AI isn't tracked yet, want it added? Request tracking.
FAQ
Is Mistral AI's assistant an autonomous agent?+
Not for most tasks. Vibe (formerly Le Chat) is chiefly an assistant for chat, web search, document and image analysis, image generation, and coding. Its Work Mode and Code Mode are genuinely agentic for multi-step tasks, but Mistral says the agent gets the user's sign-off on a plan before starting and surfaces diffs for review, so it is a supervised agent rather than a fully autonomous one.
Are Mistral's models open source?+
Partly. Mistral releases several open-weight models under permissive licenses (primarily Apache 2.0) that can be self-hosted, alongside proprietary commercial models (such as Mistral Large and Medium) available through its La Plateforme API and enterprise deployments.
What does Mistral AI cost?+
Vibe has a free tier, Pro at $14.99/month, and Team at $24.99 per user/month, with custom Enterprise pricing. The La Plateforme API is pay-per-token (for example, smaller models around $0.10/$0.30 per million input/output tokens and flagship tiers higher), with add-ons such as OCR, code execution, and web search billed per call.
Sources
- Mistral AI homepage · accessed 2026-06-20
- Mistral AI pricing · accessed 2026-06-20
- Le Chat | Mistral AI · accessed 2026-06-20
- Vibe gets to work (Work Mode) | Mistral AI · accessed 2026-06-20
- Remote agents in Vibe, powered by Mistral Medium 3.5 | Mistral AI · accessed 2026-06-20
- Mistral AI raises 1.7B EUR Series C | Mistral AI · accessed 2026-06-20
- Mistral AI (Wikipedia, company background) · accessed 2026-06-20
Last reviewed 2026-06-20