Cheap document OCR at $2 per 1,000 pages, but Mistral now points new projects to OCR 4 instead.
Kenning's take
What it does
Mistral OCR 3 turns documents such as forms, invoices, handwritten notes and low-quality scans into markdown text, with tables rebuilt as HTML. Mistral says it beats OCR 2 on handwriting, forms and complex tables. It is available through the Mistral API and the Document AI playground in Mistral Studio, and Mistral now lists the newer OCR 4 as its replacement for new work.
- Extracts text and embedded images to markdown
- Rebuilds complex tables as HTML
- Handles handwriting and low-quality scans
- Document AI drag-and-drop playground
- 50% discount through the Batch API
- Self-hosting option for enterprises
Who it suits. Developers and companies processing large volumes of documents into text or structured data.
Key facts
- Category
- AI models
- Part of
- Le Chat (Vibe by Mistral)
- Made by
- Mistral AI
- Based in
- France
- Released
- 17 December 2025
- How to get it
- through the Mistral API (mistral-ocr-2512) and the Document AI playground in Mistral Studio; self-hosting by arrangement
- Works on
- API, Web
- In the UK
- Yes
- API price
- $2 per 1,000 pages, $3 per 1,000 annotated pages, 50% off with Batch API
FAQ
When was Mistral OCR 3 released?
Mistral announced OCR 3 on 17 December 2025.
How much does Mistral OCR 3 cost?
$2 per 1,000 pages, or $1 per 1,000 pages through the Batch API.
Has Mistral OCR 3 been replaced?
Mistral lists OCR 4 as the newer model, and says OCR 3 stays available for existing integrations.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools
Devstral 2
Mistral's open-weight coding model family, now retired from the API in favour of Mistral Medium 3.5
Mistral 3
Mistral's December 2025 family of open models, from the small Ministral 3 sizes up to Mistral Large 3
Mistral Small 4
An open Mistral model that handles chat, reasoning, coding and images in one 119B mixture-of-experts model
Mistral OCR 4.1
Mistral's current document reading model, returning text with layout boxes, block types and confidence scores