Mistral OCR settings: what the page annotation changes

Mistral OCR 4.1 can fill a JSON schema of your own in the same request: its page annotation. This board measures what that changes on form pages: the values it corrects, the time and price it adds, and what one word in its prompt does.

Key takeaways

• The page annotation corrected none of 275 graded values on 15 form pages: it reads Mistral's own OCR text, never the page image.
• It makes scanned pages 2 to 3.5 times slower (a scanned W-2 took 19.4 s with it, 5.4 s without) and costs $5 per 1,000 pages instead of $4.
• One word in its prompt matters: asking for “every table” instead of “every markdown table” gave a fee table a form's 18 fields in 3 of 3 runs.

The scoreboard

How it was measured

Corrections: the 15 form pages of our forms test set (pages 13 to 15 are NIST Special Database 6 tax forms in synthetic hand print), each read once by Mistral OCR 4.1 from the page image with the page annotation on. A graded value of three or more characters counts where its exact text appears: in the OCR text, in the annotation, in both or in neither. Speed: one request per page, one at a time, medians of two runs, Sep 23 2026. Prompt wording: the closing-instructions page of a Closing Disclosure, nine runs, Oct 2 2026. Price, and what the annotation reads: Mistral's model page and annotations docs, read Oct 3 2026.

Log