Mistral OCR
Document OCR API rating Arabic strong; self-hostable; US$4 per 1,000 pages
Visit website →About
Mistral OCR is the French company's document-understanding API: it takes PDFs and images and returns Markdown with headings, tables, equations and embedded images in reading order, with an annotation mode that extracts structured fields against a schema. Mistral's language page places Arabic, Hebrew and Persian in the "strong" tier for OCR, and the service can be run as a single self-hosted container, on Amazon SageMaker or Microsoft Foundry, which is what makes it relevant to Saudi organisations that cannot send pages abroad.
Pricing in November 2026 is US$4 per thousand pages, US$2 in batch mode, and US$5 with structured annotation. Mistral's own benchmark tables across four releases include Russian, Hindi, Chinese and Turkish rows but no Arabic row; the only independent Arabic numbers come from two 2025 academic papers, one reporting a character error rate of 0.051 on synthetic print and another 0.210 on fully vocalised pages. Handwriting is claimed in general terms without an Arabic figure.
AI-assisted draft, human-reviewed before publishing — see how we choose & review tools.
Why we picked it
It is the one commercial document OCR that both names Arabic as a strong language and ships as a container you can run inside the Kingdom. We list it for that combination, and note plainly that the vendor has never published an Arabic benchmark row, so the independent papers are all a buyer has.