Paris-based AI lab founded in 2023 by former DeepMind and Meta researchers. As of May 2026 the lineup spans frontier-class flagships (Mistral Large 3, Medium 3.5) and permissively licensed open weights (Large 3, Small 4, Ministral 3 family) all under Apache 2.0.
EU-incorporated, EU-datacenter hosted, and governed under French law. For any European organization under GDPR Article 44 or the EU AI Act, Mistral is the default credible alternative to U.S.-based providers.
The April 2026 release of Mistral 3 was the biggest open-weight drop of the year. Mistral Large 3 is now Apache 2.0, ships as a 41B-active / 675B-total mixture-of-experts model, and arrives alongside Ministral 3 edge models at 3B, 8B, and 14B parameters. Mistral Medium 3.5 (v26.04) and Devstral 2 (v25.12) round out the agentic coding stack.
Recent developments
- May 13, 2026: le Chat consumer pricing migrated to USD and added a Team tier at $24.99/mo per seat (30GB storage per user, domain verification, data export). Pro held at $14.99/mo; Enterprise remains custom.
- May 12, 2026: Mistral published an SDK supply-chain advisory tied to the TanStack incident. Mistral says its infrastructure was not compromised, but compromised npm and PyPI package versions were briefly published. Teams using Mistral SDKs should check lockfiles, images, caches, and deployed artifacts for affected versions and rotate exposed secrets where needed.
- April 28, 2026: Mistral 3 shipped with Mistral Large 3 and new Ministral edge models. The Apache 2.0 release made Large 3 a 41B-active / 675B-total MoE flagship, added Ministral 3 at 3B/8B/14B with base, instruct, and reasoning variants, and went live across Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, OpenRouter, Fireworks, Together AI, Modal, IBM watsonx, and Unsloth AI on day one.
System Verdict
Pick Mistral if EU data sovereignty is non-negotiable or you want a frontier-tier model with open weights. The April 28 Mistral 3 release made Large 3 Apache 2.0, which removes the last remaining vendor-lock argument against Mistral for regulated workloads. Mistral Small 4 (Apache 2.0, 256K context) still unifies reasoning, vision, and agentic coding at commodity API cost.
Skip it for the deepest ecosystem. ChatGPT has the largest plugin marketplace. Gemini wins Google Workspace. Claude takes long-form reasoning. Mistral has none of those moats. Le Chat Pro at $14.99/mo is cheaper than ChatGPT Plus but the consumer UI is less polished.
Who pays which tier: for developers at published per-token rates. Forge is contact-sales for enterprise fine-tuning.
Key Facts
| Open flagship | Mistral Large 3 (April 28, 2026) · Apache 2.0 · 41B active / 675B total MoE · multimodal |
| Agentic coding workhorse | Mistral Medium 3.5 (v26.04) · frontier-class multimodal, optimized for agentic and coding tasks |
| Open multimodal | Mistral Small 4 (March 16, 2026) · Apache 2.0 · 256K context · 128-expert MoE |
| Edge / on-device | Ministral 3 (April 28, 2026) · 3B / 8B / 14B · Apache 2.0 · base, instruct, reasoning variants |
| Code specialists | Codestral (v25.08) · Devstral 2 (v25.12) · 80+ languages · function calling + JSON mode |
| Voice / TTS | Voxtral (v26.03) · 4B Ministral base · voice cloning · CC BY-NC 4.0 |
| Document AI | OCR 3 (v25.12) · document processing service |
| Math / formal proofs | Leanstral (v26.03) · Lean 4 proof-engineering agent |
| Enterprise fine-tuning | Forge (March 17, 2026) · license-fee pricing · contact sales |
| Consumer subscription | Free le Chat · Pro $14.99/mo · Team $24.99/mo per seat (new May 2026) · Enterprise custom |
| Data posture | EU-incorporated · EU datacenter option · GDPR Article 44 compliant · French jurisdiction |
| Open-weight lineup | Mistral 7B · Mixtral 8x7B · Mistral 8x22B · Small 4 · Large 3 · Ministral 3 family |
| Distribution | Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, OpenRouter, Fireworks, Together AI, Modal, IBM watsonx, Unsloth AI; OpenAI-API compatible |
Every data point verified on 2026-05-17 against mistral.ai/pricing, docs.mistral.ai, and Mistral security advisories.
What it actually is
Two product surfaces on one model family. la Plateforme is the commercial API for developers. le Chat is the consumer product at chat.mistral.ai.
The April 28, 2026 Mistral 3 release reshaped the open-weight tier. Mistral Large 3 is now Apache 2.0, a 41B-active / 675B-total mixture-of-experts model, and the first MoE flagship from Mistral since the Mixtral line. The Ministral 3 family ships alongside it at 3B, 8B, and 14B parameter sizes, each with base, instruct, and reasoning variants plus image understanding and multilingual support. Both families went live on day one across Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, OpenRouter, Fireworks, Together AI, Modal, IBM watsonx, and Unsloth AI.
Small 4 (March 16, 2026) still merges three prior specialized models into a single Apache 2.0 release: Magistral for reasoning, Pixtral for multimodal vision, Devstral for agentic coding. 256K context, 128-expert Mixture-of-Experts, 40% lower latency is the open-weights multilingual speech model (CC BY-NC 4.0, not permissive enough for most commercial products).
The moat is now jurisdiction plus permissive licensing on the flagship. EU-incorporated, EU-datacenter, French-law-governed. That pairing matters when GDPR Article 44 third-country transfer rules kick in, which they do for almost every European enterprise workload above a minimal threshold. With Large 3 under Apache 2.0, on-prem deployment is a procurement option for the whole frontier, not just the small models.
When to pick Mistral
- EU enterprise with GDPR or EU AI Act requirements. EU-incorporated, EU-datacenter, French-law jurisdiction. No other frontier-class provider matches this posture.
- Open weights at frontier scale. Mistral Large 3 is now Apache 2.0. Self-host the flagship behind your firewall, fine-tune for your domain, or deploy to a regulated region without contractual gymnastics.
- Edge or on-device inference. Ministral 3 at 3B/8B/14B with reasoning variants is the strongest small-model lineup in 2026 outside of Llama and Qwen.
- Multimodal reasoning at commodity cost. Small 4 handles text + vision + coding at $0.15/M input, below DeepSeek V3.2’s $0.28/M.
- Agentic coding without GitHub Copilot lock-in. Mistral Medium 3.5 and Devstral 2 cover the same workflows that Codex and Claude Code handle, with permissive weights underneath.
- OpenAI-compatible migration path.-compatible endpoint. Existing SDK integrations swap in with a base-URL change.
- Multilingual TTS for internal tools. Voxtral TTS is open-weight and multilingual. CC BY-NC 4.0 limits commercial use but internal tools are clear.
When to pick something else
- Largest ecosystem and plugin marketplace: ChatGPT. GPT Store has no Mistral equivalent.
- Google Workspace integration: Gemini. Native Docs / Sheets / Gmail hooks.
- Strongest long-form reasoning and writing: Claude. Opus 4.7 beats Large 3 on most reasoning benchmarks.
- Cheapest per-token API with open weights: DeepSeek V3.2 at $0.28/M input for text-only workloads where vision is not needed. Mistral Small 4 is cheaper but flips the trade-off toward multimodal.
- Video generation: Mistral has none. Veo 3 via Gemini or Kling.
- Commercial TTS with permissive license: Voxtral is CC BY-NC 4.0. For commercial products use ElevenLabs or OpenAI TTS.
Pricing
Consumer pricing via mistral.ai/pricing. API pricing via docs.mistral.ai.
| Plan | Price | Who’s it for |
|---|---|---|
| le Chat Free | $0 | Casual use · base models · 500 memories · document upload · image gen · connectors |
| le Chat Pro | $14.99/mo | Individuals · unlimited Large 3 · web search · image gen · extended thinking · 15GB storage · 1,000 projects · Mistral Vibe |
| le Chat Team (new May 2026) | $24.99/mo per seat | Small teams · workspace · domain verification · 30GB per user · data export |
| le Chat Enterprise | Custom | SSO, audit, compliance · contact sales |
| la Plateforme (API) | Pay-per-token | Developers building on the Mistral API |
| Forge (enterprise fine-tuning) | License fee · contact sales | Organizations training custom models on proprietary data |
per-token pricing (per 1M tokens, verify before procurement):
| Model | Input | Output | Notes |
|---|---|---|---|
| Mistral Large 3 | Apache 2.0 self-host or hosted (rates vary by provider) | Same | New April 28, 2026 · 41B active / 675B total MoE · multimodal |
| Mistral Medium 3.5 | Hosted on la Plateforme | Same | v26.04 · frontier-class multimodal · agentic and coding |
| Mistral Small 4 | $0.15 | Verify at docs.mistral.ai | March 2026 · 256K context · multimodal · Apache 2.0 |
| Ministral 3 (3B / 8B / 14B) | Apache 2.0 self-host or hosted | Same | New April 28, 2026 · base, instruct, reasoning variants |
| Codestral | Hosted on la Plateforme | Same | Code specialist · 80+ languages |
| Devstral 2 | Hosted on la Plateforme | Same | v25.12 · frontier code agent · software engineering tasks |
Prices verified 2026-05-13 via mistral.ai/pricing and docs.mistral.ai. Mistral 3 family is Apache 2.0 so hosted per-token rates vary by platform (Mistral AI Studio, Bedrock, Azure, Fireworks, Together, OpenRouter); benchmark against your preferred provider before committing.
Against the alternatives
| Mistral Large 3 | OpenAI frontier models | Claude Opus 4.7 | |
|---|---|---|---|
| License | Apache 2.0 (open weights) | Closed | Closed |
| Parameter count | 41B active / 675B total MoE | Undisclosed | Undisclosed |
| Context window | Up to 256K (Small 4) · Large 3 model card on Hugging Face | Undisclosed | 1M |
| Open weights available | Yes (Large 3, Small 4, Ministral 3, 7B, Mixtral family) | No | No |
| Edge model lineup | Ministral 3 at 3B / 8B / 14B | No | No |
| EU data residency | Native (EU-incorporated, EU datacenters) | Contractual only | Contractual only |
| Plugin ecosystem | Limited | Largest (GPT Store) | MCP-based |
| Best viewed as | EU-native open-weight frontier | Generalist default | Reasoning specialist |
Failure modes
- Hosted Mistral 3 prices vary by platform. Large 3 and Ministral 3 are Apache 2.0, so per-token rates depend on whether you use Mistral AI Studio, Bedrock, Azure Foundry, Fireworks, Together AI, OpenRouter, or self-host. Benchmark before committing.
- Output-token pricing for Small 4 not fully public. $0.15/M input is confirmed; output rate varies by source. Verify at docs.mistral.ai before budgeting.
- Voxtral license is CC BY-NC 4.0. Non-commercial only. Using it in a commercial product (even internally) requires a commercial license from Mistral.
- Smaller plugin ecosystem. la Plateforme is OpenAI-compatible, so most tools work, but Mistral-first integrations are thin compared to ChatGPT or Claude.
- le Chat UX trails competitors. Functional but less polished than ChatGPT, Claude.ai, or Gemini web. Code interpreter, canvas, and GPT-Store-equivalent features are limited or absent.
- Lineup churn is real. Medium 3.1 (v25.08) and Medium 3.5 (v26.04) overlap with Large 3 and Small 4 for adjacent workloads. Read the model docs before locking in a model ID for a long-running deployment.
- Smaller capital base than U.S. frontier labs. ~$1B raised vs tens of billions at OpenAI, Anthropic, and Google DeepMind. Compute ceiling is a real constraint on frontier model cadence.
- No native video generation. For multimodal output including video, workflow needs a second tool.
Methodology
This page was produced by the aipedia.wiki editorial pipeline, an automated system that ingests vendor documentation, verifies pricing and model details against primary sources, and generates the editorial analysis you are reading. No individual human wrote this review. Scoring follows the four-dimension rubric at /about/scoring/ (Utility, Value, Moat, Longevity, unweighted average). Last verified 2026-05-13 against mistral.ai/pricing, docs.mistral.ai/models/overview, chat.mistral.ai, and the Mistral 3 launch post.
FAQ
Is Mistral free? Yes, in two ways. The open-weight models (Mistral 7B, Mixtral 8x7B, 8x22B, Small 4, Large 3, Ministral 3 at 3B/8B/14B) are free to self-host under Apache 2.0. chat.mistral.ai also has a free consumer tier with 500 memories, document upload, image generation, and connectors.
What is Mistral 3? The April 28, 2026 open-model release. Mistral Large 3 is the new flagship: a 41B-active / 675B-total mixture-of-experts model under Apache 2.0, the first MoE from Mistral since the Mixtral line. The Ministral 3 family ships alongside at 3B, 8B, and 14B parameter sizes with base, instruct, and reasoning variants, image understanding, and multilingual support. All available across Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, OpenRouter, Fireworks, Together AI, Modal, IBM watsonx, and Unsloth AI on day one.
What is Mistral Small 4? A single Apache 2.0 model released March 16, 2026 that merges reasoning (Magistral), multimodal vision (Pixtral), and agentic coding (Devstral). 256K context, 128-expert Mixture-of-Experts, 40% faster latency and 3x higher throughput than Small 3. Available via API, Hugging Face, AI Studio, and NVIDIA NIM/NeMo.
What about Mistral Medium 3.5? v26.04 frontier-class multimodal model optimized for agentic and coding workloads. Sits between Small 4 and Large 3 in the hosted lineup. Hosted on la Plateforme, not currently open-weight.
How does Mistral compare to OpenAI frontier models? Mistral Large 3 is broadly competitive on text benchmarks and now ships under Apache 2.0. OpenAI frontier models win on ecosystem, image generation, and multimodal breadth. For EU-regulated workloads Mistral is the default pick regardless of benchmark numbers because GDPR Article 44 third-country transfer rules apply, and now you can self-host the flagship instead of negotiating a closed contract.
Why choose Mistral over OpenAI or Anthropic? Three concrete reasons: open weights at the frontier, EU-native data residency for GDPR compliance, and a credible edge model lineup via Ministral 3. U.S. developers without compliance constraints should weigh those against OpenAI’s ecosystem or Anthropic’s reasoning lead.
What is Forge? Mistral’s enterprise fine-tuning platform launched March 17, 2026. Enables custom-model training on proprietary data under a license-fee model. Contact Mistral sales for specifics.
Is Voxtral safe for commercial products? Not under the default license. CC BY-NC 4.0 restricts commercial use. Internal tools are clear; consumer products require a paid commercial license from Mistral.
Sources
- Mistral AI pricing: consumer Free, Pro, Team, Enterprise plans
- Mistral AI Studio pricing docs: per-model API rates
- Mistral model overview: current production lineup
- Mistral 3 launch post: Large 3 and Ministral 3 announcement
- le Chat: consumer product
- Hugging Face · mistralai: open-weight model releases
Related
- Category: AI Chatbots
- Comparisons: ChatGPT vs Mistral AI · Claude vs Mistral AI · DeepSeek vs Mistral AI · Gemini vs Mistral AI · Mistral AI vs Qwen
