Mistral OCR 4: good because it shows its uncertainty
- Mistral
- Automation
- Verification
On 23 June, Mistral released OCR 4, text recognition for documents. The benchmarks are strong: in an evaluation by independent annotators, OCR 4 was preferred over competing systems on average 72 percent of the time across more than 600 multilingual documents, plus top scores on OlmOCRBench (85.20) and OmniDocBench (93.07). For day-to-day work, though, something else stands out: OCR 4 returns not just text but bounding boxes, typed blocks, and a confidence score per page and per word. That is what makes the output checkable.
What OCR 4 does
It recognises text with its position, classifies blocks (titles, tables, equations, signatures), and outputs structured markdown, across 170 languages, from PDF, DOC, PPT, and OpenDocument. On the API it runs as mistral-ocr-latest; pricing is 4 dollars per 1,000 pages, 2 in batch, 5 via Document AI. Available through Mistral Studio and the console, Amazon SageMaker, Microsoft Foundry, soon Snowflake; for companies with data-sovereignty needs there is self-hosting. The output is tuned for RAG, that is citation-ready content, and one customer reports roughly four times faster processing per page than their previous provider.
Why the confidence is the real feature
OCR errors are quiet. A misread digit in an invoice or an audit-relevant record slips unnoticed into the system of record and only surfaces when it gets expensive. Confidence scores turn that around: you can route the uncertain spots to a human instead of trusting everything blindly. That is the difference between a demo and a system you can run in production. I would set the threshold deliberately and send everything below it for review, especially where the numbers matter.
Self-host when the data is sensitive
Documents are often the most sensitive thing a company holds: invoices, contracts, audit records. That OCR 4 can be self-hosted is therefore more than a checkbox. It keeps the data in house and makes you independent of the question of who still offers the service tomorrow, a point I made elsewhere in more detail.
The kind of AI that earns its place
OCR and document AI solve a real, unglamorous problem, cleanly. Use the capability. Above all, use the confidence scores. And keep a human at the point where one misread character costs money or trust.