Mistral OCR 4 tops document AI tests
Mistral's OCR 4 reads documents in 170 languages and returns structured data with confidence scores, at $4 per 1,000 pages, with self-hosting available.
Mistral released OCR 4 on June 23, a document model that reads text in 170 languages and now returns structure with it: bounding boxes, block labels for titles, tables, equations and signatures, and a confidence score on each piece. API pricing is $4 per 1,000 pages, halved to $2 on the batch API, while the fuller Document AI service is $5 per 1,000 pages.
That structured output is the real change from OCR 3, which mostly turned pages into text. Knowing where a number sits on an invoice, that a block is a table rather than a heading, and how sure the model is about each field is what lets you wire OCR into an automated pipeline without a person reading every page. Mistral reports that independent annotators preferred OCR 4 to every other OCR and document-AI system it tested, with win rates averaging 72%, and that it posted the top score on the public OlmOCRBench at 85.20.
You can call it through Mistral’s API, Amazon SageMaker and Microsoft Foundry, with Snowflake coming, or self-host the whole model in a single container if your documents cannot leave your environment.
Why this matters
If you process invoices, contracts or forms, this is cheap enough to test on a real backlog this week. Treat the confidence scores as a routing signal: auto-accept the high-confidence fields and send the rest to a person, rather than trusting the whole page. And the single-container self-host option matters for anyone whose client data cannot sit on a vendor’s servers.