On June 23, 2026, Mistral AI, a burgeoning French AI startup, officially rolled out its newest document content recognition model—OCR 4. This innovative model is capable of handling 170 languages spanning 10 language families, surpassing its rivals, including GPT 5.5 Pro and Gemini 3.1 Pro Preview. It achieved an impressive score of 93.07 on the OmniDocBench evaluation, producing results that are more in line with human preferences.
As a compact and specialized model, OCR 4 offers a range of functionalities. In addition to text output, it provides bounding box localization (which precisely locates the position of text within an image), region classification (categorizing different areas of the document), and confidence scoring (indicating the reliability of the recognition results). These features empower downstream tasks such as RAG (Retrieval-Augmented Generation) semantic chunking (breaking down text into meaningful segments based on semantics) and agent-based structuring (organizing information in a structured way using intelligent agents).
In terms of pricing, basic API calls start at $4 per thousand pages. There is a 50% discount available for bulk processing, making it even more cost-effective for large-scale operations. For document AI services, the price is set at $5 per thousand pages.
With its extensive multilingual support, outstanding performance, and budget-friendly pricing, the OCR 4 model is well-positioned to propel the development of intelligent and human-centric document processing technologies.