Mistral OCR 3
by Mistral AI
High-fidelity document understanding service with structured Markdown/HTML output, optimized for enterprise-scale document processing at $1-2 per 1,000 pages.
Overview
Mistral OCR 3 (mistral-ocr-2512) is Mistral AI's third-generation OCR service for extracting text and embedded images from PDFs and images while preserving document structure. Released December 2025, it replaces the deprecated OCR v25.03 and OCR 2.
The model outputs LLM-ready Markdown and reconstructs tables as HTML (including complex structures with rowspan/colspan), making it well-suited for RAG pipelines and downstream parsing. It demonstrates a 74% win rate over Mistral OCR 2 on forms, scanned documents, complex tables, and handwriting.
Available via API (model ID: mistral-ocr-2512) and the Document AI Playground in Mistral AI Studio. Pricing is $2 per 1,000 pages standard, $1 per 1,000 pages via Batch API. Self-hosting is available for organizations with strict data privacy requirements, though typically requires enterprise engagement.
Strengths
- Outputs LLM-ready Markdown and structured HTML tables (rowspan/colspan preserved)
- Strong performance on cursive handwriting, forms, and low-quality scans
- 74% win rate over Mistral OCR 2 on complex documents
- Industry-leading pricing ($1-2 per 1,000 pages)
- Extracts embedded images alongside text
- Doc-as-prompt capabilities for structured JSON extraction
Limitations
- Not open-source (closed weights, proprietary model)
- Self-hosting requires enterprise engagement (not download-and-run)
- SaaS-first: typical usage sends documents to Mistral's API
- Structured extraction annotations cost more than basic OCR
Best Use Cases
- RAG pipeline document parsing
- Invoice, receipt, and form extraction
- High-volume archival digitization
- Handwritten and historical document processing
- Enterprise document search and indexing