Mistral OCR 4.1 Can Now Point at the Exact Paragraph It Read

Mistral’s OCR 4.1 made the Hacker News front page on Thursday, which is unusual for a document reader. OCR stands for optical character recognition, the old and unglamorous job of turning a scan or a PDF into text a computer can work with. The new version adds three things that sound technical and turn out to matter a lot: paragraph-level bounding boxes, structural block labels, and confidence scores per block.

In plain terms: when the model reads a page, it no longer just hands back a wall of text. It tells you where on the page each paragraph came from, as coordinates. It labels what each block is, across thirteen categories including title, list, table, image, equation, and code. And it attaches a number saying how sure it is about each block, at page, block, or word level depending on what you ask for.

That combination is what makes an answer checkable. If you point an AI tool at a 200-page contract and ask what the notice period is, the useful version of that answer is not just “three months”. It is “three months”, plus a highlight on page 143 showing exactly the sentence it came from. Bounding boxes are what let a tool draw that highlight. Confidence scores are what let an automated pipeline flag the pages a human should look at instead of quietly guessing. OCR 4.1 has been in public preview since July 16 and costs €3.50 per 1,000 pages, or €4.38 per 1,000 pages with structured annotations, through the same /v1/ocr endpoint, with batch processing supported.

What is actually going on here

Document reading is where a lot of real AI work actually happens, well away from the chatbot headlines. Invoices, contracts, insurance claims, scientific papers, government forms: all of it arrives as PDFs and scans, and none of it is useful to a model until someone turns it into structured text. The reason this release leans so hard on citations and confidence is that the market has learned the hard way that an AI answer nobody can verify is an AI answer nobody will sign off on. Mistral is a French company, which matters for readers here in a practical way: European data residency is a normal option rather than a special request. Two honest caveats: this is still labelled public preview, and pricing is per page, so a large archive adds up quickly. A hundred thousand pages is €350.

What this means for you: if you have a shelf of scanned paperwork and have wondered whether AI can finally do something with it, this is the layer that makes that work. You probably will not use the API directly, but the document tools you do use are built on services like this one, and now you know what to ask them: can it show me where the answer came from? If it cannot, be more careful with what it tells you.

Sources

• OCR 4.1 model card (Mistral AI)
• OCR Processor documentation (Mistral AI)
• Mistral OCR 4: SOTA OCR for Document Intelligence (Mistral AI)