Independent annotators prefer Mistral OCR 4 over every leading OCR and document-AI system tested, with a 72% average win rate, and it hits 85.20 on OlmOCRBench, the top score among all models evaluated.
That number comes from Mistral AI's June 23, 2026 release of OCR 4, a small, focused model that finally gives production pipelines something they've been screaming for: bounding boxes, block classification, and per-word confidence scores alongside extracted text. No more guessing where a chunk of text lives on the page or whether the model is hallucinating a table.
Bounding boxes, block types, and confidence scores for every region
OCR 4 returns a structured representation per document: each block gets a bounding box, a typed label (title, table, equation, signature, and more), and inline confidence scores at the page and word level. Downstream systems use these directly instead of heuristically re-segmenting raw text.
This structure feeds semantic chunking for RAG pipelines - clean classified blocks become better retrieval units. Agents can actually act on documents (form filling, invoice processing, compliance checks) because they know where each element sits and what role it plays. The model also integrates as the ingestion component of Mistral Search Toolkit, announced at the AI Now Summit, providing citation-ready structured output for enterprise search and retrieval workflows.
170 languages, single-container deployment, $4 per 1,000 pages
OCR 4 supports 170 languages across 10 language groups, with measurable gains on rare and low-resource languages where several competing systems degrade. The model is compact enough to run in a single container, fully self-hosted, keeping document data inside your infrastructure for residency and sovereignty requirements.
API pricing is $4 per 1,000 pages; a Batch-API discount cuts that to $2 per 1,000 pages. Document AI, Mistral's no-code application-level wrapper around the same engine, costs $5 per 1,000 pages. For production at scale, those numbers compound fast - Rogo's AI engineer reports equivalent accuracy on a financial QA dataset at roughly 8x lower cost and 17x lower latency compared to leading agentic document parsers.
How the benchmarks hold up (and where they don't)
Mistral benchmarked against AI-native OCR models, frontier general-purpose models, enterprise document services, and their own OCR 3. Automated tests: OlmOCRBench 85.20 (top), OmniDocBench 93.07, internal Crawl Multilingual .98. They also ran a head-to-head human evaluation on 600+ documents across 12+ languages, sourced from third-party vendors, where annotators blindly ranked each competitor's output and preferred OCR 4 across every system tested.
Mistral acknowledges that both OlmOCRBench and OmniDocBench have known scoring limitations - annotation and formatting noise that automated metrics miss. The human preference evaluation sidesteps much of that noise by relying on judgment on realistic documents rather than string comparisons against fixed references.
OCR 4 accepts PDF, DOC, PPT, and OpenDocument formats. Teams integrate via API or through Document AI in Mistral Studio. For anyone running high-volume document ingestion, the combination of structured output, multilingual coverage, and self-hosted deployment makes this the first OCR release in years worth a serious look.
Source: Introducing Mistral OCR 4
Domain: mistral.ai
Comments load interactively on the live page.