Last updated: 2026-10-10
What changed
Mistral Large 4, in public preview since 6 October 2026, is the newest flagship from French lab Mistral AI: an open-weight Mixture-of-Experts model built for a 1 million token context window. The lab also runs Vibe, formerly Le Chat, and the OCR 4.1 document service.
About Mistral AI
Mistral AI is a French AI lab founded in April 2023 in Paris by Arthur Mensch, Guillaume Lample and Timothee Lacroix, three researchers who previously worked on large language models at DeepMind and Meta. Its September 2026 Series D valued it at over EUR 21 billion post-money, up from the $13.7 billion set by the ASML-led Series C a year earlier; the Mistral AI company profile carries the funding detail.
Mistral's strategy is to release frontier-class models as open weights alongside a hosted API and a consumer assistant. The newest flagship is Mistral Large 4, a public preview announced on 6 October 2026: a Mixture-of-Experts model with 1.05 trillion total and 52 billion active parameters, a 1.6 billion parameter vision encoder and a 1 million token context window. Mistral Large 3 remains on the price list, and document work runs through Mistral OCR 4. Mistral Small 4 folds the former reasoning (Magistral), vision (Pixtral) and coding (Devstral) lines into one configurable model.
Developers pick Mistral for EU-hosted infrastructure, per-token API prices well below the US frontier labs, and the option to self-host open-weight models for full data control. Teams building agents use the Agents API and Studio connectors, which support the Model Context Protocol for tools such as GitHub, Slack and Snowflake. Consumers use Vibe (the chat app formerly called Le Chat) on web, iOS and Android, with a CLI and IDE extension for coding; the Free plan includes Mistral Studio access.
Where the alternatives win: ChatGPT has the larger plugin ecosystem and a more polished consumer product, Claude is the stronger pick when multi-file coding reliability matters most, and Gemini fits teams living in Google Workspace. If you only want a finished assistant and have no interest in self-hosting or EU data residency, Mistral's main advantages do not apply to you. DeepSeek is the closest rival on open weights and low API cost.
Pricing
Vibe plans: Free is $0 with limited messages, web searches and coding sessions. 99. 99 per user per month.
Enterprise is custom-priced. 90. Cached input is about a tenth of the input rate, and batch jobs are half price.
Mistral OCR 4 is billed per page rather than per token. ai on 2026-10-10.
| Tier | Monthly price | What it includes |
|---|---|---|
| Free | Free | Limited messages and web searches, Limited coding sessions, Basic image generation and editing, 100+ connectors, Mistral Studio access |
| Pro | $14.99/mo | More messages and web searches, All-day coding in the CLI, IDE or on web, Spreadsheets and slides, $25.50 per month in API credits, Chat and email support |
| Team | $24.99/mo | Per-user pricing, Higher limits and faster responses, Up to 30GB storage per user, Domain name verification, Data export |
| Enterprise | Custom | Custom models, agents and workflows, Audit logs and SAML SSO, White label, Private or on-premises deployment |
Feature Comparison by Tier
| Feature | Free | Pro | Team | Enterprise |
|---|---|---|---|---|
| Price per month | $0 | $14.99 | $24.99 per user | Custom |
| Coding agent (CLI, IDE, web) | Limited sessions | All-day use | Higher limits | Custom agents and workflows |
| Bundled API credits | — | $25.50 per month | Not listed | Not listed |
| Storage per user | Not listed | Not listed | Up to 30GB | Not listed |
| Audit logs and SAML SSO | — | — | — | ✓ |
| Private or on-premises deployment | — | — | — | ✓ |
Key Features
- Mistral Large 4: Open-weight Mixture-of-Experts model in public preview since 6 October 2026, that accepts images through a built-in vision encoder and reads very long prompts.
- Vibe assistant and coding agent: Vibe, formerly Le Chat, runs on web and mobile, and the Vibe CLI and IDE extension handle all-day coding, with Work Mode executing multi-step tasks across 100+ connectors.
- MCP connectors in Studio: Studio and the Agents API support Model Context Protocol connectors for GitHub, Slack and Snowflake, so models call external systems without custom integration code.
- Low per-token API pricing: Per-token prices sit well below the US frontier labs, batch jobs are half price, and cached input costs roughly a tenth of the standard rate.
- Self-hostable open weights: Most flagship models ship as open weights, so teams can run them on their own GPUs, though commercial deployments of some models need a Mistral licence.
- OCR 4.1 document intelligence: The OCR service returns paragraph-level bounding boxes, structural block labels and block-level confidence scores, billed at $4 per 1,000 pages.
Pros
- Per-token API prices undercut the closed US frontier APIs, and batch processing halves them again, which matters for high-volume agent workloads.
- Open-weight flagships (Large 3 and the Large 4 preview) can be self-hosted, which removes per-token fees and keeps data on your own infrastructure.
- EU-hosted infrastructure with SOC 2 Type II, ISO 27001 and ISO 27701 certifications simplifies GDPR compliance for European buyers.
- Vibe Pro now bundles monthly API credits, so one subscription covers both the chat app and light API experiments.
Cons
- Mistral is an engine more than a finished product; using the raw API means building your own interface, integrations and analytics.
- Large 4 is a public preview, so its behaviour, price and open-weight licence terms can still change before general availability.
- Zero Data Retention is restricted to the Scale plan and stateless API calls; it does not cover Vibe, agents, conversations, libraries or batch files.
- The consumer plans do not include unlimited usage: Free caps messages, web searches and coding sessions, and Pro and Team raise those caps as multiples rather than removing them.
Product Information
- Cloud
- Yes
- Self-Hosted
- Yes
- On-Premise
- Yes
- Training
- Documentation
Data Handling
- Training-data policy
- Trains on input and output data by default unless Zero Data Retention is activated (Scale plan, stateless API only) or the specific product is opted out of training by default
- Data retention
- 30 days
- Compliance
- SOC 2 Type II · ISO 27001 · ISO 27701 · GDPR
Frequently Asked Questions
How much does Mistral AI cost per month?
Vibe, the consumer assistant, has a Free plan, Pro at $14.99 per month (students $5.99), Team at $24.99 per user per month and custom-priced Enterprise. Pro lists $25.50 per month of API credits. Developers pay separately per million tokens: Mistral Large 3 is $0.50 input and $1.50 output, Medium 3.5 is $1.50 and $7.50, Small 4 is $0.15 and $0.60, and the Large 4 preview is $0.68 and $2.09 at a launch sale price.
Is there a free version of Mistral AI?
Yes. The Free plan gives limited access to Vibe, including chat, search, image creation and coding in the terminal, plus access to Mistral Studio, and asks for no card. Message, search and coding-session limits are lower than on Pro. Open-weight models also run at no licence cost on your own GPUs, subject to each model's licence.
What is new in Mistral AI this month?
Mistral Large 4 entered public preview on 6 October 2026 as an open-weight flagship. Mistral raised a Series D in September 2026 at over EUR 21 billion post-money, Vibe Pro now bundles API credits, and Large 3 API pricing was cut sharply from the earlier $2 input and $6 output.
Mistral AI or ChatGPT: which should you pick?
Pick Mistral if you want open weights you can self-host, EU data residency or low per-token API costs. Pick ChatGPT for broader third-party integrations and a more refined consumer app. Mistral Pro costs less than a $20 ChatGPT plan, though ChatGPT leads on breadth. Teams comparing coding agents should also look at Claude.
Can you self-host Mistral AI models?
Yes, many Mistral models ship as open weights and can be self-hosted anywhere. Weights released under Apache 2.0 can be used freely, but some commercial deployments need a Mistral licence with separate terms for derivatives and production use. Large 4 runs only on multi-GPU nodes given its 1.05 trillion parameters, so budget for serious hardware or use the hosted API.
Top Alternatives
- ChatGPT: Pick ChatGPT when you want a ready-made assistant with a large plugin and app ecosystem and no infrastructure decisions.
- Claude: Choose Claude when long-document analysis and agentic coding quality matter more than price per token or hosting location.
- Gemini: Pick Gemini when your team already works inside Google Workspace and wants the assistant built into those apps.
HokAI guides covering Mistral AI
- Best AI Chatbots in 2026: Pick by the Job, Not the Leaderboard: ChatGPT, Claude, Gemini, Mistral's Vibe, Poe and Character.ai compared on price and actual job fit, not raw model rank, with a decision table and 2026 pricing.
- What Mistral Large 4 Is Actually For: Mistral Large 4 is a 1.05T open-weight preview with a halved sale price. Who it suits, what it costs, and what is still unknown about weights and licence.
- Best AI Assistant Apps for Android in 2026, Now That Google Assistant Is Dead: Google Assistant shut down on Android September 4, 2026. Here is how Gemini, ChatGPT, Perplexity, Copilot, Meta AI and Claude compare as its replacement.