Mistral ships OCR 4 with bounding boxes and confidence scores across 170 languages

Mistral AI launched OCR 4, targeting enterprise back-office document workflows rather than consumer chat. The model produces structured output — bounding boxes, block classification and inline confidence scores — across 170 languages, making it suitable for feeding RAG, agentic and enterprise-search pipelines rather than just dumping raw text. It's self-hostable in a single container and priced at $4 per 1,000 pages for API use, a deliberately cheap, deployable positioning.

The confidence-score and bounding-box output is the differentiator: downstream systems can decide which extractions to trust and where on the page they came from, which matters for compliance-heavy back-office use (invoices, contracts, forms). Self-hosting in one container addresses data-residency and privacy concerns that block cloud OCR in regulated industries.

Distribution is the strategic story. OCR 4 ships not only via Mistral's own API but through Amazon SageMaker and Microsoft Foundry, and Microsoft publicly flagged the launch as a milestone in its Mistral partnership — positioning the model in front of enterprise buyers already inside Microsoft's cloud. That embed-where-the-customer-already-is strategy is how Mistral extends beyond chatbots. The company is reportedly in funding talks at a €20 billion valuation while expanding European datacenter capacity, and CEO Arthur Mensch amplified the launch on X.

Competitively, OCR is a crowded space (Google Document AI, AWS Textract, Azure Document Intelligence, plus open models), so the wedge is price, multilingual breadth, structured confidence output and self-hosting together. Caveats: real-world OCR accuracy varies wildly by document quality and language, and the 170-language and benchmark claims are vendor-stated. Watch for independent accuracy comparisons and whether the SageMaker/Foundry distribution actually drives enterprise pull.