Mistral OCR 4: beats rivals in 72% blind tests, $4 per 1,000

Mistral AI released OCR 4 on Jun 24, 2026, a document-reading model that extracts text plus layout and semantic roles from files such as PDFs, Word documents, and PowerPoint slides. In a blind test covering more than 600 documents, independent reviewers preferred OCR 4's results 72 percent of the time over competing models, the company says.

What does OCR 4 do?

OCR 4 reads text and returns where each element sits on the page, plus what role it plays, and it supports 170 languages. The model produces block classifications that label page elements as titles, tables, equations, signatures, and so on, and it also outputs confidence scores for words and pages. Those features are intended to let downstream systems break documents into meaningful sections for search or to feed AI agents that need structured inputs.

OCR 4 is available through Mistral's API, Mistral Studio, and Microsoft Foundry. Mistral lists two price points: $4 per 1,000 pages for regular use, and $2 per 1,000 pages in batch mode.

How did OCR 4 perform versus competitors in tests?

In a blind evaluation with over 600 documents, independent reviewers preferred OCR 4's outputs 72 percent of the time, the company says, and Mistral adds that the model "beats all tested competitors across both benchmarks." The test size and the 72 percent preference are the specific figures Mistral provided; the company framed the result as a blind comparison against other models.

Beyond the headline number, Mistral highlights two practical strengths: the model's layout-aware block classification and the confidence scores for words and pages. Those outputs change the raw-OCR use case from simple transcription to producing structured elements that downstream systems can index or act on.

Why it matters

Layout-aware extraction and semantic block labels reduce the manual work needed to turn documents into searchable or agent-ready data. For enterprises that process contracts, tables, or mixed-format reports, a single model that returns element roles plus confidence could cut the engineering overhead of post-processing OCR output. Mistral's support for 170 languages broadens applicability to multilingual archives, while the listed price points make the economics explicit for high-volume workflows.

The 72 percent blind-test preference is a strong claim in marketing terms, and Mistral presents it as evidence of quality across both benchmarks the company ran. The practical question for customers will be whether the model's layout labeling and confidence scoring yield measurable savings in downstream tasks such as data extraction, search indexing, or agent workflows.

What to watch

Watch for independent benchmark results and third-party evaluations that reproduce the blind-test setup or probe OCR 4 on industry-specific document sets, and monitor whether competing OCR providers update models or pricing in response to Mistral's claim.