Mistral shipped OCR 4 on Monday, and the pitch isn't just cleaner text extraction. The model now returns a structured map of each document: bounding boxes around every block, typed classification for titles, tables, equations and signatures, plus per-page and per-word confidence scores. Mistral laid out the details on its research blog.
That structure is the point. Bounding boxes were the most-requested feature, and they feed the downstream work people actually care about: RAG chunking, enterprise search, redactions, and human-in-the-loop verification where someone needs to see where a flagged value sits on the page. Coverage spans 170 languages across 10 groups, with Mistral claiming the biggest gains on rare and low-resource scripts.
On numbers, treat them as company-reported. Mistral says independent annotators preferred OCR 4 over every rival tested, averaging a 72% win rate across a set of 600-plus documents in 12-plus languages, and that it scored 85.20 on OlmOCRBench. The blog itself flags that benchmark scoring has known artifacts, so the team calls the aggregate "directional rather than definitive." Fair enough.
The model is compact enough to run in a single container, so document data can stay inside a company's own infrastructure. API pricing is $4 per 1,000 pages, halved to $2 through the Batch API. The fuller Document AI layer, which reshapes output into custom JSON schemas, runs $5 per 1,000 pages. A production webinar is set for July 7.
Bottom Line
OCR 4 outputs bounding boxes, typed blocks, and confidence scores, and runs self-hosted in one container at $4 per 1,000 pages.
Quick Facts
- Price: $4 per 1,000 pages via API, $2 via Batch API
- Document AI layer: $5 per 1,000 pages
- OlmOCRBench score: 85.20 (company-reported)
- 72% average win rate in blind human evaluation (company-reported, 600+ docs)
- 170 languages across 10 language groups
- Released June 23, 2026
