Mistral OCR 4: document automation you can self-host

Every business drowns in documents someone has to type into a system: invoices, receipts, purchase orders, signed forms. Mistral OCR 4, shipped June 23, is the first document-AI release we've seen that fixes the two things that actually block automation here — knowing where text sits on the page, and knowing how sure the model is — and it runs in a single container on your own hardware. For an operator, that combination is the difference between a demo and a system you can trust with real paperwork.

What actually happened

OCR 4 returns more than a wall of extracted text. Per Mistral's announcement, it outputs bounding boxes, typed-block classification (titles, tables, equations, signatures, and more), and inline confidence scores — per page and per word — across 170 languages. It's compact enough to deploy on a single container, so regulated shops can run it on their own infrastructure for data residency.

Pricing is $4 per 1,000 pages via API, dropping to $2 per 1,000 pages on the batch API. In blind human-preference tests, annotators preferred OCR 4 over every rival system tested, with win rates averaging 72%, and it took the top score on OlmOCRBench (85.20). Mistral, to its credit, flags that these benchmarks have known scoring limits and tells you to evaluate on your own documents — good advice.

Why it matters for your business

The confidence score is the quiet unlock. At scale, you route low-confidence regions to a human and auto-approve the high-confidence ones — human-in-the-loop verification without a person reading every page. That's exactly the stop condition a document-processing automation needs to be trustworthy instead of just fast. A receipt pipeline that flags the 8% it's unsure about and clears the rest is a system you can actually put in production.

And because it self-hosts, your documents don't leave your walls, and the $2-per-1,000-pages math stays yours — no metered API bleeding on every invoice you scan. At that price, an automation that eats a few hours of manual data entry a week pays for itself almost immediately. The catch, as always: the OCR is the easy 20%. The value is in the routing rules, the exception handling, and the wiring into your accounting or ERP system — the part that turns extracted text into a booked transaction.

Sources: Mistral AI, VentureBeat.