Mistral OCR 4 : SOTA OCR for Document Intelligence

Mistral OCR 4 introduces structured document parsing, returning extracted text alongside bounding boxes, block classifications (titles, tables, equations), and inline confidence scores. The model supports 170 languages across 10 groups and operates in a single container for self-hosted deployments, enabling data privacy and cost-efficient processing. It is designed for enterprise search, retrieval-augmented generation (RAG), and agentic workflows, with structured outputs suitable for custom pipelines or no-code applications.

Independent evaluations show Mistral OCR 4 outperforms leading OCR systems, with human annotators preferring its output in 72% of cases on average. The model achieves the highest score on OlmOCRBench (85.20) and leads internal benchmarks, including Crawl Multilingual (.98) and OmniDocBench (93.07). Benchmarks have known limitations, such as ground-truth errors and scoring artifacts, which may misrepresent real-world performance.

Mistral OCR 4 integrates with Mistral Search Toolkit, an open-source framework for enterprise search and RAG workflows. Its structured outputs provide citation-ready inputs for ingestion, retrieval, and evaluation, supporting semantic chunking for RAG and structural primitives for agentic tasks like form processing. The model handles common enterprise formats, including PDF, DOC, PPT, and OpenDocument, and is optimized for high-throughput batch processing.

Pricing for Mistral OCR 4 via API is $4 per 1,000 pages, with a 50% discount for Batch API reducing the cost to $2 per 1,000 pages. Document AI is priced at $5 per 1,000 pages. Self-hosting options are available for enterprise customers, ensuring data residency and compliance. Developers can access the model via API, while teams can use Document AI in Mistral Studio for a no-code integration path.