Mistral OCR 4 Arrives, Converts PDFs and Tables Into Structured Data
Mistral OCR 4's structured output with bounding boxes and confidence scores directly addresses the bottleneck of manual formatting and human review in enterprise document workflows.
Reporting from 1 sources: GIGAZINE.
Mistral AI released Mistral OCR 4 on June 24, 2026, a document reading model that extracts text from PDFs and Office documents while identifying elements such as tables, mathematical formulas, and signatures. The model outputs structured information including bounding boxes for character and block positions, block classifications (title, table, formula, signature), and confidence scores per page and per word. In the public benchmark OlmOCRBench and Mistral AI's internal Crawl Multilingual evaluation, Mistral OCR 4 achieved the highest scores among compared models. A blind comparison by independent annotators found Mistral OCR 4's output chosen over outputs from AWS Textract, Azure Doc Intel, and Gemini 3.1 Pro Preview. The model supports 170 languages across 10 groups, with improvements in Japanese, Hindi, Greek, and low-resource languages. Pricing is $4 per 1000 pages via API, $2 per 1000 pages with Batch API, and $5 per 1000 pages for the Document AI overlay. It is available through API, Mistral Studio, Amazon SageMaker, and Microsoft Foundry, with a self-hosted option for enterprise customers who cannot send confidential documents externally.
Mistral AI positions Mistral OCR 4 for use cases including document analysis, retrieval-augmented generation, form input and invoice processing by AI agents, compliance checks, and building internal search and knowledge bases. The model supports common enterprise formats such as PDF, DOC, PPT, and OpenDocument. Mistral AI distinguishes between two usage methods: "If you need raw extraction results, use OCR 4 directly via API; if you need structured output tailored to specific business items, add Document AI functionality." The Document AI overlay aligns output to a defined JSON schema, annotates images, or interprets documents with custom instructions. In addition to API, Mistral Studio, Amazon SageMaker, and Microsoft Foundry, Mistral OCR 4 is planned to support Snowflake Parse Document. For organizations that cannot send confidential documents externally, a self-hosted option running on their own infrastructure is available for enterprise customers. The model aims to decompose documents into a form that AI and search systems can easily handle, going beyond conventional OCR that converts documents to plain text.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.