Mistral Le Chat and the Current Model Lineup: 2026 Breakdown

What Is Le Chat?

Le Chat is Mistral AI's general-purpose generative AI service. It launched in beta on February 26, 2024, with iOS and Android apps following on February 6, 2025. The name is French for "the cat." It is available on the web and mobile only. Under the EU AI Act, Mistral documents Le Chat as not high-risk.

The defining trait is dynamic routing. Le Chat does not expose a single model the way some assistants do. It sends requests to Mistral Large 3 for flagship-grade work, to Mistral Small 4 for fast general tasks, and to Mistral Medium 3 inside Le Chat Enterprise. A prototype called Mistral Next handles concise replies. Specialist models join in when a task calls for them: Voxtral for voice, Magistral for step-by-step reasoning, Mistral OCR 3 for documents, and Black Forest Labs Flux Pro for image generation.

The practical effect for a reader: you talk to one assistant, but the answer you get may come from any of several engines tuned for that job. The next section breaks down each model in that stack.

The Current Model Lineup

Two models anchor the lineup, and they are easy to confuse. Mistral Large 3 is the current flagship, released December 2, 2025. Mistral Small 4 is the newest release, from March 16, 2026, and it is the first Mistral model to fold reasoning, vision, and coding into one. Large 3 is the heavyweight. Small 4 is the versatile workhorse. Neither is a replacement for the other. You can browse the rest of our coverage from the Mistral AI hub.

Mistral Large 3 (the flagship)

Large 3 is a sparse Mixture-of-Experts model (a sparse MoE only switches on a fraction of its total parameters per token, here 41B of 675B) with 675 billion total parameters and 41 billion active per token, a 256,000-token context window, and a native vision encoder. Mistral trained it from scratch on 3,000 NVIDIA H200 GPUs and shipped it under Apache 2.0. At debut it ranked as the #2 open-weight non-reasoning model on LMArena, and Mistral announced a reasoning version was coming. API pricing is $0.50 per million input tokens and $1.50 per million output tokens.

Mistral Small 4 (the newest)

Small 4 is a 119-billion-parameter MoE with 128 experts and 4 active per token (roughly 6-8 billion active). It is the first Mistral model to unify Magistral-style reasoning, Pixtral vision, and Devstral coding in a single model, with a configurable reasoning_effort setting that ranges from low-latency replies to deep chain-of-thought. Context is 256,000 tokens, license is Apache 2.0, and API pricing is $0.20 input / $0.60 output per million tokens.

The specialists

Around those two sit the focused models. Magistral 1.2 is the reasoning line, built for traceable multi-step thinking in legal, compliance, scientific, and software work. Devstral 2 targets agentic software engineering and pairs with the terminal-native Mistral Vibe CLI. Codestral 25.08 handles instant code completion and multi-file editing across 80-plus languages. Ministral 3 covers small edge deployments. Voxtral runs voice. Mistral OCR 3 turns scanned PDFs into structured markdown or JSON.

What Le Chat Can Do

Le Chat has grown well past plain text chat. Web search returns answers with in-line citations, and a global partnership with Agence France-Presse (signed January 16, 2025) grounds responses in verified newswire content across six languages. Image generation and editing run in-chat on Black Forest Labs Flux Pro, added November 19, 2024. Voice mode runs on Mistral's Voxtral audio model. Code interpreter gives sandboxed code execution on both free and paid tiers.

For heavier work, Deep Research (preview) produces structured, source-backed reports from exhaustive web analysis. The in-chat Agents builder lets you configure autonomous agents with guardrails, custom tone, and tool access including web search, image generation, code execution, Gmail, and Google Calendar, and it is available on the Free tier. Memories (beta) gives transparent, user-controlled recall: you can see what was referenced, turn memory off, delete entries, or import from another assistant. For a step-by-step walkthrough of these features in practice, see our guide on how to use Mistral.

Le Chat Pricing Tiers

Le Chat runs four tiers. Free is the entry point at $0 per month. Pro is $14.99 per month, with a verified-student rate of $6.99 per month for up to 12 months. Team is $24.99 per user per month, or $19.99 per user per month billed annually. Enterprise is quote-based.

If you call the models through the API instead of Le Chat, pricing is per token. The two models with published per-token rates are the flagship and the newest. Our Mistral API guide shows how to authenticate and make that first call:

For a deeper plan-by-plan breakdown, see the dedicated Mistral pricing guide.

How the Models Perform

Mistral Large 3 posts 85.5% on MMLU (a 57-subject academic knowledge test, 8 languages), around 92% pass@1 on HumanEval (a 164-problem Python code generation benchmark), and roughly 43.9% on GPQA Diamond (a PhD-level expert reasoning test where ~50% is considered strong). Its SimpleQA factual-accuracy score sits near 23.8% (a one-shot factual recall test where lower means more hallucination risk), which signals real hallucination risk on fact-checking. On LMArena (April 2026) Large 3 holds roughly 1428 general Elo (a head-to-head ranking score; frontier models sit ~1500) and 1450 coding Elo. Small 4 sits near 1410 general, 1435 coding, and 1250 vision Elo.

Elo is not a zero-based scale; bars are scaled from a 1200 baseline to make gaps visible.

The reading is consistent: Mistral lands in the elite tier but below the frontier closed models. On general chat it sits comparable to GPT-4o-latest (around 1427 Elo). The gap widens on hard reasoning. Gemini 3 Pro reaches roughly 91.9% on GPQA Diamond against Large 3's roughly 43.9%, because Large 3 is a generalist, not a reasoning specialist. On coding, Large 3 stays competitive at around 1450 Elo.

Limitations and Honest Caveats

Four constraints matter most before you commit a workflow to Mistral: usage caps that apply even on the paid Pro tier, an open-weight license that is not fully open-source, the Microsoft/Azure data-sovereignty asterisk for EU residency, and reasoning and fact-check gaps against frontier closed models.

Who Should Use Le Chat?

Frequently Asked Questions