We use cookies to enhance your browsing experience, analyze site traffic, and personalize content. By clicking "Accept All", you consent to our use of cookies. You can customize your preferences or reject non-essential cookies.
Learn more about our cookie policyThe world's best document extraction engine just got a major upgrade — now with bounding boxes, typed-block classification, inline confidence scores and 170-language coverage. Sharper document AI on AI·Collab, still at 4 credits per page.
Around 90% of the world's organizational knowledge lives in documents — PDFs, scans, handwritten notes, complex technical papers. Mistral OCR is the engine that turns those documents into AI-ready knowledge on AI·Collab, and it just moved to its most capable version yet: Mistral OCR 4. Where earlier generations focused on converting a page into clean text and tables, OCR 4 returns a structured representation of the whole document. Every block is localized with a bounding box, classified by type, and given inline confidence scores — so downstream systems know not just what the document says, but where each element sits, what role it plays, and how confident the model is.
On AI·Collab, OCR 4 powers document upload and the RAG pipeline behind the scenes. You keep the same simple workflow — upload a PDF, ask questions with any of 300+ models — but retrieval is sharper because the extracted content is cleaner, better structured, and citation-ready. And the price for you is unchanged: 4 credits per page.
OCR 4 is a small, focused model that goes well beyond plain text extraction. Independent annotators preferred it over every leading OCR and document-AI system Mistral tested, with win rates averaging 72%. Here are the headline changes.
Mistral's most-requested capability: every extracted block is localized with a bounding box, so text can be highlighted in context and mapped back to its exact position on the page — the foundation for reliable, auditable data pipelines.
OCR 4 classifies each block by type — titles, tables, equations, signatures and more — turning a flat page into clean, labelled units that make better retrieval chunks for RAG and structured connectors.
Per-page and per-word confidence scores travel with the output, enabling source-grounded citations, redactions and human-in-the-loop verification where it matters — legal, financial and compliance work.
Support for 170 languages across 10 language groups, with measurable gains on specialized and low-resource scripts where many competing systems degrade sharply.
Mistral OCR 4 raises the bar on quality, but your cost on AI·Collab does not change. Document processing stays at a flat, transparent 4 credits per page — no surprise upgrade fee, no separate tier.
Credits never expire and can be used across all AI·Collab features. Usage is tracked transparently in your dashboard, so you always know exactly what you're paying for. Hard limits per document remain 1,000 pages and 50 MB.
OCR 4 handles the vast majority of document types found in organizations and everyday settings — and it keeps everything that made previous versions excellent.
Understands interleaved imagery, mathematical expressions, tables and advanced layouts such as LaTeX formatting — perfect for scientific papers with charts, graphs, equations and figures.
Reconstructs table structures with headers, merged cells, multi-row blocks and column hierarchies, preserving layout so tables stay queryable rather than collapsing into unreadable text.
Accurately interprets cursive, mixed-content annotations and handwritten text layered over printed forms — ideal for digitizing historical documents, forms and personal notes.
Processes complex formulas and equations reliably, emitting clean LaTeX — perfect for scientific papers, technical documentation and educational materials.
Accepts PDF, DOC, PPT and OpenDocument, and handles JBIG2-compressed PDFs — a common scanned format that causes crashes and extraction failures in many other systems.
Mistral evaluated OCR 4 against leading AI-native OCR models, frontier general-purpose models, enterprise document services and its own OCR 3 — in blind human preference tests and on public benchmarks.
Mistral notes that public OCR benchmarks have known scoring limitations (ground-truth errors, equivalent LaTeX marked as mismatches, multi-column reading order), which more often penalize correct output than reward incorrect output — so treat aggregate scores as directional and evaluate on your own documents.
Benchmarks, methodology and screenshots: Mistral — Introducing OCR 4 (official announcement)
The move from clean text to structured, classified, confidence-scored output changes what you can build on top of your documents in AI·Collab.
Clean, typed blocks make better retrieval units, so hybrid search finds the right passages more often — and answers cite the right part of the right page. It's the ingestion layer behind AI·Collab's RAG pipeline.
Structural primitives (blocks, boxes, types) let agents move from reading documents to acting on them — form filling, invoice processing and compliance checks — especially in legal, financial services and healthcare.
Confidence scores and bounding boxes support redactions, source-grounded citations and human-in-the-loop verification, so regulated teams can trust and audit what the model extracted.
Turn archives, technical literature, invoices and filings into indexed, answer-ready knowledge — then query it across documents with any of 300+ models on AI·Collab.
Nothing changes in your workflow — OCR 4 is already the engine behind document upload and the knowledge base. Here's the flow.
Upload a PDF or image in the chat, or add it to a knowledge base. OCR 4 automatically detects and processes the document — text, images, tables and structure.
Processing typically completes in seconds, even for complex multi-page documents. The extracted, structured content is added to your knowledge base, ready for search.
Click the file title to open document details. Keep Using Focused Retrieval on (default) for cost-effective segmented retrieval, or turn it off to load the entire document for maximum context.
Query your document with any of 300+ AI models via hybrid RAG search — including EU-hosted, ZDR-labeled options for GDPR-sensitive work.
Using Entire Document gives the most complete context but consumes far more tokens (and credits) than focused retrieval — for example, a 1,000-page document might use 288K+ tokens in full-context mode versus 3–5K with focused retrieval. Choose the mode that fits your task and budget.
No. OCR 4 is already the engine behind document upload and the knowledge base on AI·Collab. Just upload a PDF or image as usual — you automatically get the upgraded extraction quality.
No. Document processing on AI·Collab stays at a flat 4 credits per page (≈ €3.64 per 1,000 pages), regardless of the underlying model version. Credits never expire and are tracked transparently in your dashboard.
OCR 4 adds bounding boxes, typed-block classification (titles, tables, equations, signatures and more) and inline per-page and per-word confidence scores, plus expanded coverage to 170 languages across 10 language groups. In Mistral's blind human evaluation it was preferred over every competing system tested, with win rates averaging 72%.
Cleaner, structured, classified output makes better retrieval chunks, so the RAG pipeline surfaces the right passages more reliably and answers cite the right part of the right page. It's especially noticeable on complex documents — tables, equations, multi-column layouts and multilingual content.
Up to 1,000 pages and 50 MB per document. Multiple documents can be added to the same knowledge base for cross-document search — for example two 800-page books become 1,600 pages searchable together.
Documents are processed for extraction and stored in your knowledge base for retrieval. You can query them with EU-hosted, ZDR-labeled models for GDPR-sensitive work — see the model catalog for the current ZDR/EU-hosting status per model.
Benchmark figures and capability descriptions are from Mistral's published OCR 4 announcement (mistral.ai/news/ocr-4) and reflect the state at publish time. Pricing reflects the AI·Collab model catalog. Credit rates, limits and model labels can change — the model catalog and your account dashboard are always the authoritative source for live prices and availability.
Transform documents into AI-ready knowledge with Mistral OCR. Process up to 1000 pages at just 4 credits per page with 94.89% accuracy.
Read moreLearn how AI·Collab transforms PDFs into accurate AI answers — from OCR to embedding, hybrid search, and EU-hosted reranking. All data stays in Europe.
Read moreWhat “context” means on our model cards, why it matters, and how to work efficiently with long or short context windows.
Read moreGet started today. Access models from OpenAI, Google, Anthropic, Grok and more.
GDPR compliant · Zero data retention · Cancel anytime