Mistral OCR

Mistral OCR 3 is a specialized document intelligence model from Mistral AI that extracts text and embedded images from a wide variety of documents with exceptional fidelity. It outputs clean markdown enriched with HTML-based table reconstruction, preserving both content and structure for downstream systems. The model handles challenging document types such as handwriting, complex tables, low-quality scans, and forms, outperforming both traditional enterprise OCR solutions and other AI-native approaches. Key features include a breakthrough 74% overall win rate over Mistral OCR 2 in internal benchmarks, state-of-the-art performance on forms, handwriting, and tables, and an industry-leading price of $2 per 1,000 pages (with a 50% Batch-API discount reducing it to $1 per 1,000 pages). The model is significantly smaller than competitive solutions, enabling faster and cheaper inference without sacrificing accuracy. Mistral OCR 3 is available via API (model ID: mistral-ocr-2512) and through the Document AI Playground in Mistral AI Studio, which offers a drag-and-drop interface for parsing PDFs or images into text or structured JSON. It also offers a self-hosting option for organizations with stringent data privacy requirements. Designed for developers, data extraction pipelines, and enterprise document workflows, Mistral OCR 3 balances accuracy and cost. However, Mistral has since released OCR 4, which supersedes OCR 3 with further improvements. Users evaluating new deployments should consider the latest version.