Mistral released OCR 4 on 23 June 2026, the latest version of its document-intelligence model — and the headline change is that it no longer just turns a page into text. Alongside the extracted content, OCR 4 returns bounding boxes that locate each element on the page, typed-block classification (titles, tables, equations, signatures and more), and inline confidence scores generated per page and per word. The model supports 170 languages across 10 language groups, runs in a single container for fully self-hosted deployment, and is positioned as an ingestion layer for retrieval-augmented generation (RAG) and enterprise search.

What is new versus OCR 3

Earlier generations of Mistral's OCR focused on converting a page into clean text and tables. OCR 4 returns a structured representation instead: for each block, downstream systems get not only what the document says but where the element sits, what role it plays, and how confident the model is in each region.

That extra structure is what makes the upgrade more than a version bump. Bounding boxes — which Mistral describes as its most-requested capability — allow text to be highlighted in context and matched reliably back to the source, which matters for source-grounded citations and for data pipelines that cannot tolerate silent misalignment. Block types and confidence scores feed redactions, human-in-the-loop verification, and cleaner semantic chunking for RAG, where well-classified blocks make better retrieval units. For agentic workflows, the structured output gives an agent the primitives to act on a document — filling a form, processing an invoice, running a compliance check — rather than just reading it. OCR 4 accepts common enterprise formats including PDF, DOC, PPT and OpenDocument.

Multilingual coverage and self-hosting

The language coverage is the other substantive claim. OCR 4 supports 170 languages across 10 language groups, and Mistral says its internal multilingual evaluation showed OCR 4 leading across the eight groups it tested, including Southeast Asian and specialised low-resource language groups. Treat that as a vendor benchmark, not an independent language-coverage audit, but it is still the right feature to evaluate if your document estate spans multiple scripts.

Equally relevant for many buyers is deployment. Mistral says OCR 4 is compact enough to run in a single container, which means an organisation can self-host it and keep document data inside its own infrastructure for residency, sovereignty and compliance reasons. Self-managed deployment is offered to enterprise customers; the model also slots into Mistral's open-source Search Toolkit (in public preview) as the ingestion stage of a RAG or enterprise-search pipeline.

What it costs and where to get it

Through the API, OCR 4 is priced at $4 per 1,000 pages, dropping to $2 per 1,000 pages with the 50% Batch-API discount. The no-code Document AI path in Mistral Studio — which wraps the same engine — is $5 per 1,000 pages. A single API endpoint always returns the extracted content, bounding boxes, block types, confidence scores and markdown-structured text; the Document AI parameters layer structured-JSON output, image annotation and custom-prompt interpretation on top of the same call when you need the result reshaped to a schema. Availability spans the Mistral API and Studio, Amazon SageMaker and Microsoft Foundry, with Snowflake support listed as coming soon, plus the self-hosting option for stricter data-privacy requirements.

The benchmark caveat worth reading

Mistral reports strong numbers, but the benchmark caveat is essential. The company says independent annotators preferred OCR 4 over the competing systems it tested with average win rates of 72%, and it reports scores of 85.20 on OlmOCRBench and 93.07 on OmniDocBench. Mistral also says the public benchmarks have scoring limitations and that competitor scores are its own internal reproductions, so the results should guide an evaluation rather than settle one.

For anyone weighing a switch, that is the right framing: vendor-run benchmarks comparing a vendor's own model against the vendor's reproductions of competitors are a starting point, not a verdict. The model also comes with a scope limit Mistral states plainly — OCR 4 is a document-understanding model, not a decision-maker, and is not intended for medical diagnosis, legal judgment, high-stakes financial decisions, safety-critical systems or real-time processing.

What it means for the region

Two of OCR 4's design choices line up well with how documents actually move through regional businesses. The first is language: Southeast Asia is deeply multilingual, and the document-heavy sectors that anchor the regional economy — banking and financial services, shipping and logistics, government administration, legal services — routinely handle forms, contracts and records across multiple scripts. A model that holds accuracy on lower-resource regional languages is more useful here than a benchmark-topping English-only system.

The second is self-hosting. Mistral says OCR 4 can run in a single container for self-managed enterprise deployment, which may help organisations keep sensitive documents inside their own infrastructure and manage confidentiality, cross-border transfer and governance concerns. That does not remove the need for a proper privacy assessment, but it gives regulated buyers an option beyond sending every document to a hosted API. Neither point requires taking the benchmark claims at face value — both are structural advantages that hold regardless of where OCR 4 lands on a leaderboard.

Key Takeaways

  • Mistral OCR 4, released 23 June 2026, returns structured output — bounding boxes, typed-block classification and per-page/per-word confidence scores — rather than just extracted text and tables.

  • It supports 170 languages across 10 language groups; Mistral's internal evaluation (across the eight groups it tested, including a Southeast Asian group) reports the largest gains on specialised and low-resource languages, and Mistral says the model can run self-hosted in a single container.

  • API pricing is $4 per 1,000 pages ($2 with the Batch discount); the Document AI no-code path is $5 per 1,000 pages. It is available via Mistral's API/Studio, Amazon SageMaker and Microsoft Foundry, with Snowflake coming soon.

  • Mistral's benchmark wins (≈72% human-preference, OlmOCRBench 85.20, OmniDocBench 93.07) are vendor-run with competitor scores it reproduced internally; Mistral itself calls them directional and recommends evaluating on your own documents. OCR 4 is a document-understanding tool, not a decision-maker.