# Mistral OCR 4 Mistral OCR 4 is [[Mistral AI]]'s document understanding model, and the version where [[Mistral OCR]] stops being a text extractor and starts returning **structured documents**: bounding boxes for every element, block classification (title, table, equation, signature, …), and per-page plus per-word confidence scores. Those three additions matter more than any accuracy bump. Confidence scores let a pipeline route low-confidence pages to review instead of trusting everything blindly; bounding boxes make citations and highlighting possible; block types turn a flat text stream into something a downstream system can query. ## Facts that matter - **170 languages**, with claimed accuracy holding on low-resource ones - **Inputs**: PDF, DOC, PPT, OpenDocument - **Deployment**: Mistral Studio API, Amazon SageMaker, Microsoft Foundry, or a single self-hosted container (data-residency use cases) - **Pricing**: $4 per 1,000 pages via API ($2 batch), $5 per 1,000 pages via Document AI - **Benchmarks**: 85.20 OlmOCRBench, 93.07 OmniDocBench, 72% average win rate in independent preference evals — with Mistral itself flagging that OCR benchmarks suffer from annotation errors and math-notation mismatches. Rare vendor honesty; still, treat the numbers directionally. - Plugs into Mistral's Search Toolkit for [[Retrieval-Augmented Generation (RAG)]] ## Where it fits The paid escalation tier in a parsing pipeline: triage locally with [[pdf-inspector]] or [[liteparse]], send only scanned/complex pages here. ## References - [Announcement](https://mistral.ai/news/ocr-4/) - [Documentation](https://docs.mistral.ai/capabilities/document/) ## Related - [[Mistral OCR]] — the product line - [[Mistral AI]] - [[pdf-inspector]] · [[liteparse]] — local triage before paying per page - [[Retrieval-Augmented Generation (RAG)]]