Mistral AI releases OCR 4 for document intelligence

On June 23, 2026, Mistral AI released OCR 4, a compact document intelligence model designed for enterprise search, retrieval-augmented generation (RAG), and agentic workflows.

Unlike traditional OCR systems that produce flat text streams, OCR 4 returns a structured representation of each document. Every text block is localized with a bounding box, classified by type (title, table, equation, signature, etc.), and assigned page-level and word-level confidence scores.

The model supports 170 languages across 10 language groups and is priced at $4 per 1,000 pages, with a 50% discount for batch API usage. It is small enough to run in a single container for fully self-hosted deployments, addressing data-residency and sovereignty requirements.

In independent human evaluations, annotators preferred OCR 4 over competing systems in 72% of documents. It also achieved an OlmOCRBench score of 85.20, the highest recorded on that benchmark.