Mistral AI released OCR 4 on 23 June 2026, adding paragraph-level bounding boxes, structural block labels and confidence scores to its document extraction service. Rather than returning only a stream of text, the model records where material appeared, what role it played and how certain the extraction was.

Mistral AI Watch analysis
What happened
OCR 4 recognises blocks such as titles, tables, equations, captions, signatures, headers and footers, while generating confidence information at page and word level. Mistral says it accepts common business formats including PDF, word-processing files, presentations and OpenDocument, and covers 170 languages across ten language groups. The service is available through the model interface and a higher-level structured-document route; enterprise customers can also pursue a self-managed deployment in a single container.
The published price is $4 per 1,000 ordinary pages, or $5 per 1,000 pages when structured annotations are requested. Mistral also advertises a batch discount for the ordinary extraction route. Those figures describe the provider's service at launch, not the full cost of validation, storage, indexing or human review.
Why it matters
For retrieval systems and document agents, preserving structure can matter as much as recognising words. A table cell detached from its heading, a footnote promoted into the main argument or a signature mistaken for body text can quietly poison later answers. Bounding boxes and typed blocks give downstream systems a chance to cite the right region, route uncertain passages to a person and keep reading order intact. That is less glamorous than a talking robot, which is probably why it may prove more useful.
Our read
Mistral reports an 85.20 score on OlmOCRBench, 93.07 on OmniDocBench and a 72 percent average win rate in its own blind human-preference study. The company also says every competitor result was reproduced internally and warns that automated benchmarks can mishandle equivalent mathematics, reading order and flawed reference annotations. That candour is welcome. It also leaves the sensible test unchanged: run the invoices, scans, handwriting and dreadful legacy forms that actually live in your cupboard.
What to watch
- Independent evaluations across low-resource languages and messy scanned archives.
- Whether confidence scores reliably identify expensive extraction mistakes.
- How self-managed throughput and operating cost compare with the hosted service.
- Whether agents preserve block provenance when they transform extraction into decisions.
Discussion spark: Are document agents limited more by reasoning quality, or by whether extraction preserves tables, layout and provenance correctly?
Sources and evidence
- Introducing OCR 4 (23 June 2026, 00:00 UTC)
- OCR 4.0 model card (23 June 2026, 00:00 UTC)
not affiliated with or endorsed by Mistral AI