Mistral OCR 4.1

Mistral OCR 4.1 is a document-processing service for extracting text, structure and layout information from supplied pages.

Preserve where the text came from

The service documents paragraph bounding boxes, structural block labels and block confidence scores. These are useful when an application must connect extracted content back to a page region. Keep page identifiers and geometry alongside text instead of flattening every document into one anonymous string.

OCR and question answering are different jobs

Extraction produces the material a downstream search or answering system can use. It does not, by itself, decide whether a claim in the document is correct. For a document assistant, first inspect the extracted table, heading and reading order, then test whether the answering stage cites the right passage.

Pricing follows pages

The model page lists $4 per thousand pages and $5 per thousand annotated pages. This profile does not translate those charges into a fabricated per-million-token rate. Check the requested output features, document size and current billing rules when estimating a batch.

Build a representative test set

Include scans with rotated pages, small footnotes, multiple columns and tables whose headers continue across pages. Score missing text, incorrect numbers and the loss of relationships between rows and columns. Use documents from the actual upload path: preprocessing and compression can change extraction quality.

For a local checkpoint alternative, inspect DeepSeek OCR 2. Compare accepted pages, processing cost and the work required to repair an extraction, rather than judging only whether some text appeared.

Sources

Publisher source 1 Deployment limits and commercial terms belong to the selected checkpoint or API route.