Mistral AI released Mistral OCR 4 on June 23, 2026.
The model adds bounding boxes and block classification to extracted text. It provides inline confidence scores for every word and page.
Independent annotators prefer OCR 4 over all tested competitors. Human evaluation win rates average 72% across systems. OCR 4 scored 85.20 on the public OlmOCRBench dataset.
The tool supports PDF, DOC, PPT, and OpenDocument formats. It handles 170 languages across ten language groups.
Users can run the model via API or Document AI. The API costs $4 per 1,000 pages with a 50% batch discount. Self-hosting keeps sensitive data within your own infrastructure.
Why it matters
Run agents that act on documents with reliable source grounding. Get structured inputs for retrieval-augmented generation (RAG) pipelines. Process high-volume documents at eight times the cost of competitors.
Source: https://mistral.ai/news/ocr-4/
