Mistral OCR 4: Document AI for Business Automation

Mistral OCR 4 is a document-AI model, released June 23, 2026, that returns a structured representation of any enterprise document — bounding boxes, block types, and confidence scores — rather than the flat wall of text earlier OCR generations produced. For teams building document-automation pipelines, that shift is the headline: extraction stops being a parsing problem and becomes a structured-data problem.

It is Mistral’s fourth OCR generation in roughly 15 months, and it lands in a crowded field — Google Document AI, Amazon Textract, Azure Document Intelligence, ABBYY, and a wave of open-weight models. What sets OCR 4 apart is less any single benchmark number and more the combination of aggressive pricing ($4 per 1,000 pages, $2 in batch), a single-container self-hosting option, and a structured output that removes integration layers teams used to build by hand.

This guide covers what actually shipped, why the structured output changes automation economics, an honest read of the benchmark claims (several of which are vendor-stated), a recomputed cost comparison against the hyperscalers, and how to put OCR 4 to work without overpromising on the numbers.

Mistral describes the leap plainly. Where earlier generations focused on converting a page into clean text and tables, OCR 4 returns a structured representation of the document. In practice that means every extracted element arrives with its location on the page, a type label, and a confidence score — three things downstream automation historically had to reconstruct or guess at.

The model accepts PDF, DOC, PPT, and OpenDocument files directly, so the full spread of back-office documents flows in without a pre-conversion step. Mistral reports coverage across 170 languages in 10 language groups, with gains on rare and low-resource languages where competing systems tend to degrade — a meaningful detail for multinational document pipelines.

Bounding boxes were OCR 4’s most-requested feature, and the reason is structural. Without location data, a downstream RAG or compliance pipeline cannot trace an extracted fact back to the page it came from — the traceability gap that makes audit-ready extraction genuinely hard. With coordinates attached to every element, an extracted number can point back to the exact cell it was read from.

Block classification does similar work one layer up. OCR 4 assigns every element a type — title, table, equation, signature, and others — as a first-class model output rather than a separate post-processing stage. That removes an integration layer enterprise teams used to build in-house just to tell a heading from a table. Confidence scores complete the set: they operate at both page and word level, which is what makes confidence-gated automation possible.

There is a second, related product worth separating clearly. OCR 4 is the extraction model. Document AI is a Studio product, priced at $5 per 1,000 pages, that wraps OCR 4 with a second-pass mistral-small-2603 call to reshape the output into custom JSON schemas. If your automation needs fields in a fixed shape rather than a generic structured document, Document AI is the mode to evaluate — and its dependence on the smaller model is one reason the Mistral Small model family matters to the document stack.

Strip away the benchmarks and the durable advantage is price. OCR 4 is $4 per 1,000 pages on the standard API and $2 per 1,000 in batch mode — a 50% discount for non-interactive workloads. Schema-driven Document AI is $5 per 1,000. The table below annualizes published list pricing across three volumes; every Mistral, Azure, and Google cell is the per-1,000-page rate multiplied out by hand.

The 100K-page row is the one to sit with. A firm processing 100,000 pages a year pays $200 in OCR 4 batch mode versus $3,000 on Azure Document Intelligence’s custom extraction tier — a 15x gap that holds at every volume. At the standard $4 API rate the gap is about 7.5x. Azure’s Read tier matches OCR 4’s entry economics at $1.50 per 1,000, but it returns text without the bounding boxes and block types that make OCR 4’s output automation-ready.

A caveat on the bottom row. Baidu’s open-weight Unlimited-OCR has no per-page license fee, but “free” is not zero: you pay for GPU compute, deployment, and operations. A precise per-page figure depends on your hardware, utilization, and throughput, so the table marks those cells as an infrastructure estimate rather than a quoted rate. The honest comparison is a managed per-page price against an amortized infrastructure cost you have to model for your own load.

Mistral reports a top-line OlmOCRBench score of 85.20 and calls it the “top overall score.” That claim deserves a caveat. The public OlmOCRBench leaderboard — last updated May 21, 2026, before OCR 4’s release — places Infinity-Parser2-Pro at 87.6 and Chandra-2 at 85.9 above it, and VentureBeat independently notes OCR 4 would rank roughly third on the current public board. OCR 4’s 85.20 is a vendor-submitted figure that does not yet appear on the independently reproduced leaderboard.

The rest of the benchmark story is similarly vendor-framed, and worth keeping in that frame. OlmOCRBench itself, built by the Allen Institute for AI, runs 7,010 unit tests across 1,403 PDFs in seven categories, with per-score uncertainty of roughly a point either way — so small gaps between models are inside the noise. The figures below are Mistral’s own; treat them as directional.

What does the 85.20 actually measure? The table below maps OlmOCRBench’s seven categories to the back-office failure each one predicts — the difference between an abstract leaderboard and a decision about whether to trust extraction on your own documents.

OCR 4 ships three ways to consume it, and the third is the strategic one. Beyond the managed API and the cloud marketplaces, Mistral supports a single-container self-hosted deployment — letting a regulated enterprise process sensitive documents entirely inside its own infrastructure, with no routing to an external U.S.-jurisdiction cloud API. For organizations weighing the broader tradeoffs, our self-hosted deployment decision guide covers the infrastructure side in depth.

OCR 4 arrives against established hyperscaler document AI and a fast wave of open-weight models. The most direct counterpoint shipped one day earlier: Baidu’s Unlimited-OCR, a 3-billion-parameter, MIT-licensed model that parses entire PDFs in a single forward pass and gathered roughly 1,800 GitHub stars in its first 24 hours. It is free and self-hosted — and it has no managed API and no enterprise SLA, which is exactly the gap OCR 4’s paid tier fills. Mistral’s own open-weight model lineage is part of why a self-hosting story is even credible from this vendor.

The pricing makes more sense as strategy than as a margin play. The global intelligent document processing market was about $2.30B in 2024 and is projected to reach $12.35B by 2030 at a 33.1% CAGR, with BFSI the largest segment. OCR 4 feeds directly into Mistral’s Search Toolkit as the ingestion layer for RAG and enterprise search — so winning document extraction is really about owning the front door to every downstream AI workflow.

The financial backdrop fits that ambition. Mistral is targeting €1 billion in revenue for 2026, up from roughly €200 million in 2025, and is reportedly in early discussions to raise about €3 billion at a valuation near €20 billion — nearly double its €11.7 billion Series C from September 2025. No round has been announced as of late June. Pricing OCR 4 to undercut the hyperscalers by an order of magnitude is how you buy share in a market growing at 33% a year, and pairs naturally with Mistral’s broader enterprise AI stack.

The practical payoff of confidence scores is a pipeline that does not ask a human to read every page. Set a threshold; auto-approve regions above it; route the rest to review. Bounding boxes give the reviewer the exact spot to look, and block types let you apply different rules to tables, signatures, and free text. That is the difference between an OCR tool and a document-automation system — and it is where the customer testimonials, hedged appropriately, point.

Two of those testimonials are worth quoting with that hedge in mind. Rogo, a financial-AI firm, reported reaching equivalent accuracy at roughly 8x lower cost and 17x lower latency versus leading agentic document parsers; Anaqua, an IP-management firm, reported OCR 4 is roughly 4x faster per page than its incumbent. Both are customer statements on single, undisclosed datasets — directional evidence, not reproduced benchmarks. The right move is to run OCR 4 against your own documents before you commit a forecast to it.

If you are mapping document automation onto a real back-office process — invoice capture, claims intake, contract review, CRM data entry — the structured output is the input, but the value is in the workflow around it. That scoping work is exactly what our AI digital transformation engagements start with: a confidence-gating design and an honest per-page cost model before any vendor commitment.