カオスマップに戻る

Overview

Mistral AI is a Paris-based French AI company known for efficient open-weight models. European origin, GDPR compliance, and data sovereignty are core design principles. The general-purpose line is a three-tier ladder — Mistral Large 3 / Medium 3.5 / Small 4 — surrounded by task-specific models: Mistral OCR 4 for documents, Codestral for code, Voxtral for audio, and Shieldstral for safety classification.

Model lineup

General-purpose (Large 3 / Medium 3.5 / Small 4)

ModelBest forAPI pricing (input / output per 1M tokens)
Mistral Large 3Flagship: advanced function calling, multilingual work$0.50 / $1.50
Mistral Medium 3.5128B blended model, 256k context; open weights under a modified MIT license, self-hostable on as few as 4 GPUs$1.50 / $7.50
Mistral Small 4Volume work: classification, extraction, routine generation$0.15 / $0.60

At $0.50 per 1M input tokens, Large 3 is unusually cheap for a model claiming frontier capability. The default split is Small 4 for bulk, Medium 3.5 for long context and self-hosting, Large 3 for the hard requests.

Task-specific models

Mistral names its specialized models "-stral" and serves them from the same API as the general-purpose line.

ModelBest forAPI pricing (input / output per 1M tokens)
CodestralCode generation, completion, review (80+ languages)$0.30 / $0.90
Mistral OCR 4Structured extraction from invoices, contracts, papers, scans
Voxtral SmallAudio (transcription, TTS)$0.10 / $0.30
Pixtral LargeImage understanding, chart reading$2.00 / $6.00
Magistral MediumReasoning$2.00 / $5.00
Ministral 8BEdge and on-device$0.15 / $0.15

Additions during 2026 include Leanstral 1.5 (open-source math and code), Robostral Navigate (embodied navigation), and Shieldstral 1.0 3B, a 3B open-weight multimodal safety classifier.

Key features

1. European origin and GDPR compliance

Designed with EU data sovereignty in mind, which matters for European enterprises. In August 2026 Mistral started offering in-EU data processing and priority access, though limits on eligible workloads and plans mean not everything stays inside the EU.

2. Open weights

Many models ship as open weights. Small 4 is Apache 2.0 and Medium 3.5 uses a modified MIT license, so self-hosting carries no per-token cost.

3. La Plateforme API

An OpenAI-compatible interface that makes migration easy. Since August 2026 the platform also serves third-party open models through the same API, starting with Z.ai's GLM-5.2.

4. Le Chat / Vibe

Mistral's own chat app. It has been rebuilt as the "Vibe" experience with agent execution, custom MCP connectors, and memory; Mistral Vibe for coding is included from Pro upward.

5. Mistral Studio (prompt and skill management)

A system of record for prompts and skills with versioning, ownership, and tracing — the governance layer enterprises need once agents reach production.

Pricing and plans

Le Chat subscriptions and API usage are billed separately: paying for Pro grants no API credits.

PlanMonthlyFor
La Plateforme APIPay-per-useDevelopers
Le Chat Free$0Individuals (soft cap around 25 messages/day)
Le Chat Pro$14.99Individuals and pros (150-message soft cap; extended thinking, deep research, Mistral Vibe)
Le Chat Team$24.99/userTeams (30GB storage, admin tools)
Le Chat EnterpriseContact salesSAML SSO, audit logs

At $14.99, Le Chat Pro undercuts ChatGPT Plus and Claude Pro at $20, which makes it the obvious pick when price is the deciding factor.

Typical uses

  • LLM adoption in European companies where GDPR compliance is mandatory
  • Software development with Codestral
  • On-premises LLM via open weights
  • Document processing — turning invoices and contracts into structured text with Mistral OCR 4
  • Multilingual applications, especially French, German, and Spanish
  • Agent building on strong function calling

Strengths

  • Large 3 at $0.50 / $1.50 is exceptionally cheap for a frontier-class model
  • European, GDPR-compliant, data-sovereign
  • Many open-weight models (Apache 2.0 / modified MIT)
  • A full set of task-specific models — code, documents, audio, vision, reasoning, safety
  • Strong multilingual support, especially European languages
  • La Plateforme is OpenAI API-compatible

Weaknesses

  • ❌ Smaller ecosystem than OpenAI or Anthropic
  • ❌ Below GPT-5.6 and Claude Opus 5 at the top tier
  • Limited Japanese optimization
  • In-EU processing is limited by workload and plan, so not every feature stays within the EU

Official resources

公式サイト