Mistral releases OCR 4.1 with paragraph-level bounding boxes and confidence scores

Mistral has published OCR 4.1, a new version of the document-recognition service behind its Document AI stack. The model is listed in public preview in Mistral's documentation, where the pages describing it are dated July 16, and it appears in the model catalogue under the identifier mistral-ocr-4-1.

What is new in 4.1

The additions are structural rather than claims about raw accuracy: native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores. Together the three features describe output that arrives annotated — which text belongs to which block, where that block sits on the page, and how confident the service is in each block — rather than a flat stream of recognised characters. The confidence signal matters for pipelines that must decide when to send a page to a human reviewer instead of trusting the model's output automatically.

Inside a wider Document AI stack

Mistral positions the service inside a Document AI suite that combines OCR with structured data extraction and targets enterprise document processing across many document types and languages; a separate page lists the supported languages. Alongside OCR 4.1 the suite offers annotations for pulling structured fields out of documents, document question answering that pairs other models with OCR, and batch processing for large jobs. The model comparison page lists OCR, structured annotations, bounding box extraction and batching among the supported features.

Availability and pricing

The documentation still labels the model a public preview rather than a generally available release. Mistral's model comparison guide quotes $4 per 1,000 pages for OCR and $5 per 1,000 annotated pages, a shape that reflects metering by pages processed rather than tokens — relevant for teams estimating the cost of digitising large archives or invoice backlogs. For those weighing document-AI options, the practical question is less a single accuracy number than whether the structured output fits existing review workflows.