Overview
Mistral OCR 4 is a new-generation document recognition model released by French AI company Mistral AI on June 23-24, 2026. Its core positioning shifts from traditional "text extraction" to deeper "structured document understanding." This is Mistral's "visual frontend" for the Document AI stack—it not only reads text but also returns paragraph-level bounding boxes and structured block labels (headings, tables, lists, formulas, footnotes, etc.), turning scanned documents, PDFs, and images directly into "structured data consumable by LLMs."
**Covering 170 languages**—this is its biggest killer feature compared to general multimodal models like GPT-5.5 Pro and Gemini 3.1 Pro: stable recognition of European minority languages, Arabic, Hindi, and various Southeast Asian scripts. On benchmarks, it **outperforms mainstream multimodal models such as GPT-5.5 Pro and Gemini 3.1 Pro** (per Mistral's official leaderboard).
**Pricing is extremely aggressive**: API base call **$4 per thousand pages** (approx. ¥30), **batch processing with 50% discount** reduces to $2 per thousand pages; the upper **Document AI suite costs $5 per thousand pages** (including OCR + structured parsing + workflow)—lower than AWS Textract, Google Document AI, Azure Document Intelligence, etc. This is a cost-effective option for developers working on RAG document preprocessing, enterprise knowledge bases, contract/invoice/financial report parsing.
This is Mistral's sharpest differentiation breakthrough amid the dual pressures from "full-stack large models" and vertical AI companies—**a landmark product from European AI entering the document intelligence track.**
Key Features
- Paragraph-level Bounding Boxes: Natively outputs precise bounding boxes for each paragraph, suitable for downstream PDF reflow/highlighting/retrieval positioning
- Structured Block Labels: Distinguishes structural blocks like headings/paragraphs/tables/lists/formulas/footnotes, directly feeding LLMs without losing layout information
- Covers 170 Languages: Includes European minority languages, Arabic, Hindi, and various Southeast Asian scripts, essential for multinational enterprise document scenarios
- Outperforms GPT-5.5 Pro: Exceeds GPT-5.5 Pro / Gemini 3.1 Pro and other general multimodal models on benchmarks (official leaderboard)
- API $4 per Thousand Pages: Base API call $4/thousand pages, batch processing with 50% discount reduces to $2/thousand pages
- Document AI Suite: $5/thousand pages packages OCR + structured parsing + workflow, ready for enterprise out-of-the-box
- Native Mistral Ecosystem: Deep synergy with Le Chat / Codestral / Mistral Large, EU compliance friendly
Use Cases
- Document preprocessing for RAG knowledge bases (PDF/scanned documents/images → structured text)
- Contract/financial report/invoice/parsing
- Batch digitization of multilingual documents (Europe, Middle East, Southeast Asia)
- Layout-preserving OCR for academic papers/textbooks
- Enterprise internal document intelligent search/extraction/archiving
- Customers with strict EU data compliance needs (GDPR/AI Act)
Pros
- Paragraph-level bounding boxes + structured block labels, output directly consumable by LLMs
- 170 language coverage, essential for multinational enterprises
- Outperforms GPT-5.5 Pro / Gemini 3.1 Pro on benchmarks
- $4/thousand pages (batch $2) pricing is a strong disruptor for enterprise OCR
- Friendly to European AI companies and customers sensitive to compliance and data sovereignty
- Document AI suite out-of-the-box, from OCR to structured processing in one package
Pricing
**Mistral OCR 4 API**: $4/thousand pages; **Batch processing** 50% discount, $2/thousand pages. **Document AI Suite**: $5/thousand pages (includes OCR + structured parsing + workflow). **Self-hosted/Enterprise**: Contact Mistral sales.
Summary
Mistral OCR 4 is the most impactful OCR/Document AI release of June 2026—**paragraph-level bounding boxes + 170 languages + performance exceeding GPT-5.5 Pro + $4/thousand pages** makes it stand out cost-effectively against incumbents like AWS Textract, Google Document AI, and Azure Document Intelligence. If you're building RAG knowledge bases, parsing contracts/financial reports/invoices, or digitizing multilingual documents, Mistral OCR 4 is one of the most worthy options to evaluate; if you're already using Gemini/GPT multimodal for OCR, consider offloading that part to Mistral OCR 4 to reduce costs.
Version History
- Mistral OCR 4 (2026-06-23): Paragraph-level bounding boxes + structured block labels; 170 languages; outperforms GPT-5.5 Pro / Gemini 3.1 Pro; API $4/thousand pages (batch $2); Document AI suite $5/thousand pages