Mistral OCR 4 API Review 2026: Document Intelligence at $4/1K Pages

Mistral OCR 4, announced in early August 2026, is the latest in Mistral's line of document-intelligence models — and it is a deliberate break from "just read the text" OCR. The model (id mistral-ocr-4-0) returns bounding boxes, typed-block classification (titles, tables, equations, signatures, lists) and inline confidence scores alongside extracted text, supports 170 languages across 10 language groups, and runs in a single container for fully self-hosted deployments. It is priced at $4 per 1,000 pages via the API, or $2 per 1,000 pages with the 50% Batch-API discount.

In blind human-preference evaluations, independent annotators preferred OCR 4 over every leading OCR and document-AI system tested, with win rates averaging 72%, and it posted the top overall score on the public OlmOCRBench (85.20). That combination — segmentation-quality output, aggressive per-page pricing and self-hosting — makes OCR 4 a serious look for enterprise search, retrieval-augmented generation (RAG) and agentic ingest pipelines. This APIRank review verifies the facts, breaks down the real cost per page, and compares it against the document-AI incumbents.

📰 What Mistral OCR 4 actually is (verified 2026-08-03)

Mistral positions OCR 4 as document intelligence, not text extraction. The headline features from the official announcement:

• Segmentation, not just text. Each returned block carries a bounding box (coordinates), a typed classification (title, subtitle, text, image, table, equation, signature, thinking) and an inline confidence score. Downstream steps can reason about a table cell or a signature region instead of re-parsing flat text.
• 170 languages across 10 language groups. Practical for multilingual scanned invoices, legal filings, contracts and archival digitization.
• Single-container self-hosting. The whole model ships in one container for on-premise deployment, aimed at regulated industries where documents must not leave the building.
• Integrated with Mistral Search Toolkit (public preview). OCR 4 acts as an ingestion component, feeding structured, source-grounded output into the open-source search framework announced by Mistral.

Mistral's own customer quote — from Rogo, a financial-AI company — frames the economic case: on a chart- and figure-dense financial QA dataset, the team reached equivalent accuracy at roughly 8x lower cost and 17x lower latency than the incumbent agentic document parsers they were using.

💰 Pricing (verified from official announcement)

Mistral publishes OCR 4 pricing per page, not per token — a useful simplification for document workloads. All prices are from the official OCR 4 release notes.

What that looks like at volume. A team ingesting 10,000 pages per month would pay about $40 through the synchronous API or $20 through the Batch API — versus thousands of dollars for the same volume run through a token-billed frontier vision model, where every page can consume hundreds of thousands of image tokens. At 100,000 pages a month the Batch rate drops the cost to $200, which is why the per-page model is the headline economic argument: predictable, linear, and cheap enough to scan an entire back-catalogue archive without a budget fight.

🔧 Structured output: the real differentiator

What makes OCR 4 more than a scanner-with-text is the structured metadata it returns per block. Where a classic OCR API gives you a string of words and rough coordinates, OCR 4 gives you:

• Bounding boxes for every detected region, so layout is preserved for downstream re-flow or redaction workflows.
• Typed-block classification — titles, subtitles, text, images, tables, equations, signatures, "thinking" blocks — so a table cell can be routed to a table parser and a signature region to a KYC check without heuristics.
• Inline confidence scores per element, which drive source-grounded citations, redactions and human-in-the-loop verification.

For teams building retrieval pipelines this is the difference between "found the text near a table" and "know this is the signature block with 0.98 confidence." That is the same rationale as the structured outputs emerging across modern vision OpenAI-compatible APIs.

🧠 Benchmarks: where OCR 4 wins

Mistral cites three benchmark families, all from the official release notes:

• Human Preference Evaluations: independent annotators preferred OCR 4 over every leading OCR and document-AI system tested, averaging a 72% win rate.
• OlmOCRBench: top overall score of 85.20 among the systems tested.
• OmniDocBench: 93.07; internal Crawl Multilingual evaluation .98, ahead of both AI-native and enterprise solutions.

Mistral flags a caveat itself: automated benchmarks like OlmOCRBench and OmniDocBench have known scoring limitations (annotation and formatting noise), which is why the human-preference result is the headline. Treat raw numbers as directional, not absolute.

💼 Use cases that fit OCR 4 best

The per-page pricing and structured output point at specific workloads where OCR 4 beats both generic OCR and chat-vision:

• Enterprise RAG & knowledge bases. Ingest PDFs, scanned contracts and slide decks, keep the layout and confidence metadata, and feed clean chunks into a vector index for retrieval-augmented generation. The typed blocks let you chunk on real semantics — one page becomes a set of title, paragraph, table and formula chunks with their own coordinates and scores.
• KYC, compliance & contract review. Signature blocks and confidence scores are the raw material for automated signature detection, document redaction and human-in-the-loop verification, cutting the manual triage load in onboarding and e-discovery.
• Insurance claims & healthcare. Structured extraction of forms, medical records and policy documents, with the self-hosted container keeping protected data inside the organisation.
• Libraries & archival digitization. The 170-language coverage and single-container deployment suit national archives, newspapers and multilingual heritage collections that need batch, on-premise processing.
• Financial statements. Chart- and table-dense filings are exactly where Mistral's cited customer saw 8x lower cost and 17x lower latency than agentic parsers.

If a workload is mostly "read one image and tell me what it shows," keep using a chat-vision model — OCR 4 is a document pipeline tool, not a general image Q&A endpoint.

🆚 Mistral OCR 4 vs the document-AI incumbents

How does OCR 4 stack up against the mainstream options teams already use for document extraction?

The clearest wins for OCR 4 are (1) predictable per-page pricing with a real discount for volume, and (2) first-class self-hosting. If you are already on an OpenAI-compatible stack for chat, note OCR 4 is a dedicated OCR model, not a chat-completions model — plan for a separate endpoint call for document jobs.

🛠️ Getting started (Python SDK)

Calling OCR 4 is a single request through the standard Mistral Python SDK. A minimal example:

from mistralai import Mistral
client = Mistral(api_key="YOUR_MISTRAL_KEY")

# Point at a PDF or image; the page comes back as text + metadata
resp = client.ocr.process(
    model="mistral-ocr-4-0",
    document={"type": "document_url", "document_url": "https://example.com/contract.pdf"},
)
for page in resp.pages:
    print(page.index, page.markdown[:120])   # extracted text
    for block in page.blocks:
        print(block['type'], block.get('confidence'))  # typed blocks + scores

For high-volume jobs, prefer the Batch API: submit a manifest of document URLs once, poll for results, and pay the $2/1,000-page rate instead of $4. The Batch route also suits latency-insensitive backfills — annual report scans, archival digitization, invoice re-processing — where you want throughput and cost efficiency over per-call latency. For context on Mistral's broader model line, see our earlier Mistral AI API review.

🔐 Security & data-privacy posture

Data residency is where OCR 4 separates from token-billed frontier vision APIs. Because Mistral runs from EU data centers, documents processed through the managed API stay under EU data-protection rules by default. And because the model is shipped as a single container, teams in regulated sectors can run the entire pipeline on-premise or inside their own VPC — no document ever leaves the network. The per-block confidence scores also enable policy-driven redaction: low-confidence PII or signature regions can be flagged automatically before a human reviews them, which fits finance, legal and healthcare audit requirements. If you already run workloads on managed document AI, compare the residency story on both sides before committing.

⚠️ Limitations to know

• No China regional endpoint. Mistral serves from EU data centers with no mainland China access program; cross-border production needs a proxy or aggregator, or a domestic OCR fallback.
• Per-page pricing, not per model mix. Budget is simple, but there is no per-token granularity, so very short one-line scans still cost the same per page.
• Benchmark caveats. OlmOCRBench and OmniDocBench have known scoring artifacts; verify on your own document mix before committing.
• Dedicated OCR endpoint. It is not a general chat-vision model — structured document parsing is the focus, not open-ended image Q&A.

🎯 Verdict

Mistral OCR 4 is a strong pick for structured, high-volume document intelligence — especially if you need EU residency or self-hosting. The $4/1,000-page (or $2 batch) pricing is competitive and predictable, the structured output (boxes, block types, confidence) is genuinely useful for RAG and KYC pipelines, and the single-container deployment option is rare in this category. It is less ideal for one-off open-ended image questions (use a chat-vision model) and requires proxy plumbing from mainland China. If your workload is multilingual scanning at scale with a compliance requirement, OCR 4 is worth a serious pilot.

FAQ

What is Mistral OCR 4?

Mistral OCR 4 is a document-intelligence OCR model (id mistral-ocr-4-0) released August 2026 on the Mistral La Plateforme API. Unlike plain text OCR, it returns bounding boxes, typed-block classification and per-element confidence scores, plus self-hosted single-container deployment.

How much does Mistral OCR 4 cost?

The API costs $4 per 1,000 pages, with a 50% Batch-API discount bringing it to $2 per 1,000 pages for asynchronous high-volume jobs. The managed Document AI product is $5 per 1,000 pages. Self-hosting uses a Mistral license.

How many languages does Mistral OCR 4 support?

It supports 170 languages across 10 language groups, practical for multilingual scanned invoices, legal filings, contracts and archival digitization.

How well does Mistral OCR 4 perform on benchmarks?

Blind human-preference evaluations preferred OCR 4 over every leading OCR and document-AI system tested, averaging a 72% win rate. It scored 85.20 on OlmOCRBench, 93.07 on OmniDocBench and .98 on Mistral internal Crawl Multilingual.

Can I self-host Mistral OCR 4?

Yes. OCR 4 runs in a single container for fully self-hosted deployments, attractive for regulated industries (legal, healthcare, government, finance) needing on-premise documents.

Can Mistral OCR 4 be used from mainland China?

Mistral serves from EU data centers with no mainland China regional endpoint or official China access program. Use a stable overseas proxy or aggregator, or a domestic OCR endpoint as a latency-safe fallback for high-volume scanning.

Related reads

• Mistral AI API review — models, pricing & China access
• OpenAI-compatible API landscape
• Anyscale API review — Ray & Endpoints
• DeepSeek API review
• Qwen Image 3.0 API pricing

Sources: Mistral API Docs / OCR 4 announcement, Mistral documentation, TechInsider coverage. Pricing and benchmark figures verified 2026-08-09.