Mistral OCR

Mistral OCR is Mistral AI's document understanding API, launched March 6, 2025 as 'the world's best document understanding API.' It converts PDFs and images into structured Markdown/JSON, handling complex layouts, interleaved images, tables, mathematical expressions, and LaTeX formatting while preserving document structure. Mistral reports 94.89% accuracy versus 83.42% for Google Document AI and 89.52% for Azure OCR, with throughput up to 2,000 pages per minute and multilingual support across thousands of scripts. Pricing is aggressive at $1 per 1,000 pages (batch: $0.50), and it plugs into RAG pipelines as the extraction layer. Later versions (OCR 3, OCR 4) added bounding boxes, confidence scores, and schema-driven annotation.\n\nThe honest picture: the accuracy figures are Mistral's own benchmarks — independent verification is thin, and 'world's best' is marketing. It's API-only (with selective on-premise for enterprises), so no open weights and no offline use. Documents cap at 50MB/1,000 pages per request. The OCR 3/4 upgrades raised prices ($2-4 per 1K pages), so the famous $1/1K pricing applies to the original tier via batch. Even with those caveats, it became the default document-extraction layer for RAG stacks in 2025, which says a lot about how well the core product works.