Mistral launches OCR 3 to understand complex documents

Mistral has launched Mistral OCR 3, a model that turns difficult-to-read documents, such as forms, invoices, handwritten pages and complex tables, into text and structured data that other systems can use.

The company says OCR 3 achieves a 74% overall win rate against Mistral OCR 2 in internal tests using real-world business documents. The comparison included forms, scanned documents, complex tables and handwriting.

OCR stands for optical character recognition: technology that analyzes an image or PDF and extracts the text it contains. More advanced versions do more than read words. They also try to understand where headings, columns, checkboxes and images are located.

What improves in practice

Mistral OCR 3 is designed to work with documents that often cause traditional systems to fail:

• Forms, even when they include checkboxes, labels, dense fields or handwritten annotations.
• Handwriting, including cursive and text written over printed documents.
• Low-quality scans, with noise, skew, distortion, compression or low resolution.
• Complex tables, with merged cells, multiple header rows and column-based structures.

For tables, the model can reconstruct the structure using HTML tags such as colspan and rowspan. This preserves relationships that would be lost if the document were simply converted into a list of words.

For example, a company could upload hundreds of invoices and receive structured fields such as vendor, date, amount and taxes. It could also digitize a handwritten historical archive or turn scanned technical reports into content that a search engine or AI assistant can query.

A tool you can use without coding

Mistral has added OCR 3 to Document AI Playground, a drag-and-drop interface available in Mistral AI Studio. From there, you can upload a PDF or image and receive the content as clean text or JSON, an organized format that is useful for connecting information to other programs.

Developers can also access it through the API using the mistral-ocr-2512 model. The company says it is compatible with previous versions of Mistral OCR 2, making it easier to replace the model in existing workflows.

The pricing targets high-volume use

Mistral sets the cost at 2 dollars per 1,000 pages. When using its Batch API, designed to process jobs without an immediate response, the price drops 50% to 1 dollar per 1,000 pages.

Cost matters because extracting text is the first step in many enterprise AI systems. If a model cannot read a contract, invoice or report accurately, the assistant working with that data will also have an incomplete or incorrect foundation.

Mistral says OCR 3 outperforms OCR 2 across all evaluated languages and document formats. Its tests are based on comparisons with correct answers and internal business-use examples, so real-world performance may vary depending on the quality and type of files each organization processes.

The strategy is clear: make documents that once required manual review usable directly in search, automation and AI agent systems. The next thing to watch is how it performs beyond the test examples, especially with legal documents, sensitive data and particularly difficult handwriting.