Mistral OCR 4 returns document structure, not just text

Mistral released OCR 4, and the change from earlier versions is what it gives you alongside the text. Previous releases mostly turned a page into a stream of characters. OCR 4 returns bounding boxes for where each element sits, a type label for what it is (title, table, equation, signature), and an inline confidence score for every block. So instead of a flat transcript, you get something a downstream pipeline can route, filter, and check.

The numbers Mistral reports are strong. The model leads OlmOCRBench at 85.20 and OmniDocBench at 93.07, and in blind human preference tests independent annotators chose its output over leading competitors about 72 percent of the time. It covers 170 languages grouped into ten families, with particular attention to low-resource scripts where most OCR systems fall apart. Pricing is $4 per 1,000 pages on the API, $2 in batch mode, and $5 for the Document AI mode that emits structured JSON. There is also a single-container self-hosted build for teams that cannot send documents to an external service. Mistral is candid that the benchmarks have ground-truth errors and notation mismatches, so the scores understate real performance in places.

The confidence scores are the part worth testing first. They let you send only the uncertain blocks to a human reviewer instead of re-checking everything, which is the difference between a demo and a production document workflow.

Why it matters

If you build RAG or agent pipelines that read PDFs, invoices, or forms, per-block confidence lets you set a threshold and human-review only what falls below it. That turns OCR from a step you have to trust blindly into one you can audit, so test the confidence scores against your own hardest documents before wiring it in.