Mistral AI Releases OCR 4 — Text Recognition System with Blocks, Borders, and Confidence Scores

Mistral AI has announced Mistral OCR 4, a new optical character recognition model that not only extracts text but also returns bounding boxes, block-level classification (headings, tables, equations, captions, and more), and a confidence score for each word. The model supports 170 languages across 10 language families and can be deployed in a single container for fully local data storage. In human evaluations of over 600 documents in 12+ languages, independent assessors preferred OCR 4 over all tested OCR systems and document AI services, with an average preference rate of 72%. On the public OlmOCRBench benchmark, the model scored 85.20 points, ranking first among tested systems. On the OmniDocBench benchmark, it scored 93.07, though the company notes known limitations of automated tests: some discrepancies stem from errors in reference data, equivalent mathematical notation, equation segmentation, and reading order in multi-column documents. OCR 4 is integrated into the Mistral Search Toolkit (an open-source search framework) and is available via API ($4 per 1,000 pages, with a 50% discount for batch mode at $2) or through Document AI in Mistral Studio ($5 per 1,000 pages). Document AI adds the ability to output structured JSON according to a given schema and image annotation. The model is recommended for document parsing, RAG (Retrieval-Augmented Generation), agentic workflows, enterprise search, and invoice processing. The company emphasizes that OCR 4 is not intended for medical diagnosis, legal opinions, or safety-critical systems.