Mistral OCR 4 Explained: What’s New in Document Intelligence

Mistral OCR 4 turns document extraction into a structured perception task. Released on June 23, 2026, it returns not only text and tables but also bounding boxes, block categories, and confidence scores that tell downstream systems where content appears, what role it plays, and where review may be needed.

That structure is the main advance over a plain OCR transcript. A retrieval system can preserve page location for citations. A redaction workflow can identify signatures or headers. An invoice agent can route low-confidence fields to a human instead of treating every extracted character as equally reliable.

Mistral supports OCR 4 through its API and Document AI, with access through Mistral Studio, Amazon SageMaker, and Microsoft Foundry. Enterprise customers can also discuss self-hosting in a single container when documents must remain in their own infrastructure.

From text extraction to document structure

OCR 4 breaks a page into localized blocks. Each block has a bounding box and a type, such as title, text, table, equation, image, or signature. The response also includes markdown-structured content and confidence at page and word level.

Those outputs enable several production patterns:

• highlight the exact source region behind an answer;
• create semantic chunks based on headings and tables;
• preserve reading order across complex layouts;
• send uncertain words or fields to a verifier;
• distinguish body text from headers, footers, and signatures;
• keep equations and tables separate from prose.

A language model can reason over plain text, but it loses evidence when spatial and structural context disappears. OCR 4 keeps more of that context available to retrieval and agent systems.

Supported formats and languages

Mistral says the model accepts common enterprise formats, including PDF, DOC, PPT, and OpenDocument. It supports 170 languages across ten language groups, with particular emphasis on specialized and lower-resource languages where competing systems may degrade.

Broad coverage is not proof of equal quality in every language. Fonts, scans, vertical text, historical spelling, mixed scripts, handwriting, tables, and page damage can all change accuracy. Teams should build a validation set from their actual archive rather than testing only clean English PDFs.

OCR 4 is a focused document model. Mistral explicitly lists raw audio, video, and non-document inputs as out of scope.

API extraction versus Document AI

Mistral offers two layers around the same OCR engine.

Pure OCR extraction is appropriate when developers want the raw content, bounding boxes, block types, and confidence. The application owns schema mapping and business logic.

Document AI adds an interpretation layer. A developer can provide a JSON schema or custom instruction, and the OCR output is passed into another Mistral model to produce domain-specific structured data or image annotations.

The additional Document AI layer is convenient, but it adds another probabilistic interpretation. Preserve the underlying OCR and source location so users can verify a structured field.

Published pricing

Mistral lists OCR 4 API extraction at $4 per 1,000 pages. The published 50% Batch API discount reduces that to $2 per 1,000 pages. Document AI is priced at $5 per 1,000 pages.

Those rates make page count easy to estimate, but total system cost includes:

• file conversion and storage;
• retries and duplicate pages;
• structured post-processing;
• embeddings and retrieval;
• human verification;
• self-hosted infrastructure and operations;
• downstream model calls.

Batch pricing is valuable for archives that do not need interactive latency. A user waiting on one form may need the standard path. Keep separate queues so a million-page ingest does not block an urgent document.

Benchmarks with unusually clear caveats

Mistral reports an 85.20 score on OlmOCRBench, 93.07 on OmniDocBench, and an average 72% human-preference win rate across its comparison set. The company also explains why aggregate OCR scores can be misleading.

Reference annotations may contain mistakes. Equivalent LaTeX can be marked wrong because strings differ. Multi-column reading order and equation segmentation can change the score even when the rendered meaning is correct. Header and footer handling can create artificial mismatches.

Mistral therefore calls the benchmark numbers directional and recommends evaluation on real documents. That is the correct production posture. Compare field-level accuracy, reading order, structural fidelity, and reviewer effort on the documents that matter to the business.

Confidence scores are routing signals

Per-word confidence makes selective review possible. It is not a calibrated probability that a word is true in every domain. A system should test whether low confidence actually correlates with errors on its own data.

Useful routing rules include:

1. auto-accept high-confidence, low-impact fields;
2. compare important numbers against deterministic checks;
3. send low-confidence signatures, dates, totals, or identifiers to review;
4. preserve bounding boxes in the review UI;
5. reject the page when reading order is clearly broken;
6. record corrections for ongoing evaluation.

Never let a confidence threshold approve a medical, legal, financial, or safety-critical decision. OCR supplies evidence; it does not own the decision.

Prepare documents before extraction

Input quality still shapes the result. Detect encrypted or corrupt files, normalize page rotation, and identify duplicate pages before paying for OCR. Preserve the original file and avoid destructive image enhancement; a sharpening filter can make a page look clearer while changing a faint digit.

For large documents, record page-level success so a timeout does not force a complete restart. Validate that the returned page count matches the source, that every block belongs to a known page, and that bounding boxes stay within page coordinates. Tables, equations, and multi-column pages deserve targeted review even when their average confidence is high. These controls turn a strong parser into a dependable ingestion pipeline.

RAG and agent use cases

OCR 4 integrates with Mistral Search Toolkit, which was in public preview at launch. Structured blocks can become cleaner retrieval units, and bounding boxes can support source-grounded citations.

For agents, the model supplies primitives for form filling, invoice processing, compliance checks, and archive search. A safe agent should reference the source page and region, distinguish extraction from inference, and ask for review before writing to a system of record.

Prompt injection can exist inside a document. Extracted text that says “ignore your instructions” is content, not authority. Keep tool permissions and system policy outside the document context.

Deployment and privacy

The hosted API minimizes infrastructure work. Self-hosting keeps documents within an organization’s environment but transfers responsibility for updates, capacity, logging, vulnerability management, and model monitoring.

A single-container deployment may simplify packaging, but production scale still needs queueing, storage, observability, and failover. Compare total ownership cost with the hosted page price rather than treating self-hosting as free.

From document understanding to a finished document workflow

OCR 4 extracts and structures content. It does not replace document conversion, PDF assembly, or distribution. Those are separate operations after the extracted information has been reviewed.

For example, Claude might use OCR 4 to understand a scanned source and prepare an index. An approved DOCX report can then be converted through an external service such as Medux, and several approved PDFs can be assembled into one package. Medux is not an OCR 4 host, and OCR is not a Medux feature in this workflow.

The Claude MCP DOCX-to-PDF tutorial shows signed upload, asynchronous conversion, polling, and download. The PDF merge-and-split tutorial covers controlled document assembly.

A traceable pipeline should preserve:

• original document and checksum;
• OCR model version and page-level output;
• bounding boxes and reviewer corrections;
• approved DOCX or PDF source;
• Medux task ID for conversion or merge;
• final page count, order, and checksum.

If OCR is rerun, do not silently replace an approved PDF. If conversion fails, do not repeat OCR. Keeping understanding and file transformation as independent stages makes the workflow easier to audit and prevents one model’s output from becoming an unverified final document.