Mistral OCR 3, a precise, structured and affordable OCR

Maxime Hiez
Mistral AI
15 Jan, 2026

Introduction

In December 2025, Mistral AI announced the launch of Mistral OCR version 3, an Optical Character Recognition (OCR) API that sets a new standard for document understanding. This advanced technology enables the processing and transcription of complex documents with unparalleled accuracy and speed, delivering unprecedented document comprehension capabilities.

What’s new with OCR 3

Significantly improved form accuracy : Handwriting, degraded scans, and complex tables, with a 74% improvement compared to OCR 2 (internal evaluations on business cases, etc.).
Structured output : OCR3 produces Markdown that preserves formatting and reconstructs tables in HTML (rowspan/colspan).
Designed for automation : The API response includes JSON with a list of pages (markdown per page, images, detected links, optional headers/footers), and structured annotations.

Developer and Studio Experience

Template : mistral-ocr-2512 (integration via public API).
Document AI Playground in Mistral AI Studio : Import a PDF or image and retrieve clean text or structured JSON without coding.
Smart placeholders : Markdown references images / tables via placeholders (![img-0.jpeg], [tbl-3.html]) resolved from the image arrays / tables in the JSON — this simplifies faithful document reconstruction.

Pricing and availability

Standard : 2$ / 1 000 pages
Batch API : -50% or 1$ / 1 000 pages
Structured Annotations : 3$ / 1 000 pages

At these price points, OCR 3 outperforms many enterprise document extraction solutions while delivering excellent quality for complex cases (forms + handwriting + tables).

Use cases

Invoices, purchase orders, KYC : Ready-to-use field extraction + limited human validation thanks to HTML/JSON structures and layout preservation.
Scanned archives and files : Increased resilience to degraded scans; better management of handwritten notes and annotations to create searchable archives or train AI agents.
Complex tables (banking, healthcare, public) : HTML + Markdown reconstruction facilitates ingestion into analytical pipelines.

Why now ?

Mistral positions OCR as the first building block of an enterprise AI strategy. As long as critical data on paper/PDF is not digitized and structured, AI use cases remain confined to the proof-of-concept stage. OCR 3 aims to unlock these data resources and accelerate the transition from pilot to production (process automation, agents, etc.).

How to get started ?

Test in the Playground (PDF/Image -> text/JSON) to validate the quality of your documents.
Prototype the mistral-ocr-2512 API ; enable the Batch API for large volumes (cost ÷2).
Store the generated JSON/HTML files, connect a post-processing tool (rules, mapping, LLM, RPA), and monitor.

Conclusion

Mistral OCR3 delivers next-generation OCR : accurate on complex cases, structure-aware for automation, and cost-effective at scale. For organizations looking to open their document data to AI (agents, analytics, etc.), it’s an immediate accelerator ; simple to test, quick to deploy, and affordable.