Mistral OCR 4.1 preview adds paragraph-level bounding boxes and confidence scores

Mistral has pushed a public preview of OCR 4.1, the document-recognition engine behind its Document AI stack. The headline additions over prior versions are structural: the model now returns paragraph-level bounding boxes, labels for structural blocks, and per-block confidence scores. That combination moves it beyond flat text extraction toward layout-aware parsing, which matters for downstream tasks like table reconstruction, form processing, and feeding cleanly segmented text into retrieval or LLM pipelines.

The confidence scoring at block granularity is the most practically useful change. Rather than trusting or discarding a whole page, developers can flag or re-review only the low-confidence regions, which is how OCR realistically gets integrated into automated document workflows. Native bounding boxes also reduce the need for a separate layout-detection pass.

Pricing is set at €3.5 per 1,000 pages, rising to €4.38 per 1,000 pages when annotations are included. As a public preview rather than a stable release, the API surface and output format could still shift before general availability, so teams building against it should expect some churn.