Mistral OCR 4 review, pricing and verdict

Built for enterprise teams parsing PDFs and office documents at scale, especially regulated buyers who cannot send files to a public API. In Mistral's blind test across 12+ languages, annotators preferred it over every rival OCR system tested. It replaces manual tagging pipelines and general OCR tools, not image-understanding models.

Mistral OCR 4 is Mistral AI's document extraction model, released in June 2026, returning text, paragraph-level bounding boxes, 13 structural block labels and confidence scores across 170 languages. Mistral publishes no downloadable weights, so access goes through Mistral's API, Microsoft Foundry, Amazon SageMaker, or a self-hosted container licensed through Mistral sales.

Provider: Mistral AI · Family: Mistral OCR

More about Mistral AI on Hokai

Input modalities: image, pdf · Output: text

About Mistral OCR 4

Mistral OCR 4 is a document intelligence model from Mistral AI, released on June 23, 2026 as mistral-ocr-4-0. It is the fourth generation of a lineup that began with Mistral OCR (mistral-ocr-2503, March 6, 2025), then OCR 2 (mistral-ocr-2505, May 22, 2025) and OCR 3 (mistral-ocr-2512, December 18, 2025). A point release, OCR 4.1 (mistral-ocr-4-1), shipped on July 16, 2026 and reached general availability on August 31, 2026. According to Mistral's changelog, the mistral-ocr-latest and mistral-ocr-4 aliases have resolved to 4.1 since July 16, while mistral-ocr-4-0 stays GA for teams that pin it.

Can you download Mistral OCR 4? Not as public weights. Mistral publishes no OCR model files (its Hugging Face organization listed no OCR repository when we checked on September 29, 2026), so every deployment goes through one of four routes named in Mistral's June 23 announcement and its Azure deployment docs.

The Mistral API in Mistral Studio is self-serve: call the /v1/ocr endpoint with a Mistral API key, billed per page with a Batch API discount. On Azure, Microsoft Foundry lists it under Document AI, deployable as a serverless Model-as-a-Service endpoint billed through the Azure subscription. Amazon SageMaker carries OCR 4 and Document AI through Mistral AI's AWS Marketplace listing, billed by AWS. The self-hosted container runs inside your own infrastructure under an enterprise agreement with Mistral, and its price is not published.

The same announcement listed Snowflake Parse Document as coming soon. On Azure the service runs on Microsoft infrastructure and on SageMaker it runs on Amazon Web Services, which matters for teams whose data-residency rules name a specific cloud.

With include_blocks set to true, each page returns a blocks array in reading order, and each block carries a paragraph-level bounding box and one of 13 structural labels: text, title, list, table, image, equation, caption, code, references, aside_text, header, footer and signature. The confidence_scores_granularity parameter returns page, block or word scores, with block granularity added alongside OCR 4.1. The pages parameter accepts integer lists or range strings such as '0,2-4', and the Batch API handles large sets asynchronously. Document AI runs on the same endpoint with extra parameters: pass a JSON schema and the OCR output is fed to mistral-small-2603 to return schema-shaped JSON. Mistral lists PDF, DOC, PPT and OpenDocument among the accepted formats.

On the public OlmOCRBench, Mistral reports the top overall score among the systems it tested (85.20), plus 93.07 on OmniDocBench. Mistral notes that both benchmarks have known scoring limitations. For its human evaluation it assembled 600+ documents in 12+ languages from third-party vendors, and independent annotators ranked the outputs blind. They preferred OCR 4 over every system tested, with win rates averaging 72%. Rogo, a financial-services AI platform, is quoted in the announcement as matching leading agentic document parsers on a financial QA set at roughly 8x lower cost and 17x lower latency.

Mistral's trust center lists ISO 27001, ISO 27701 and SOC 2 Type II (checked September 29, 2026). Mistral's announcement says OCR 4 is not intended for medical diagnosis, legal judgment or high-stakes financial decisions made without human review.

Where the alternatives win: OCR 4 is a narrow model. It does not caption photos or answer questions about charts, so general image work belongs on a vision-language model such as Mistral Medium 3, and reasoning over text you have already extracted fits Mistral Large 3. Teams that need files they can run offline today, without an enterprise contract, should look at open-weight document models such as North Micro Vision Instruct from Cohere or LFM2.5-VL-3B from Liquid AI. Buyers who want a hosted parsing product with its own interface rather than a raw model endpoint can compare Reducto and the vendors in our LlamaParse vs Reducto vs NuMind comparison. More document-capable models are in the vision model directory.

Pricing

$4 per 1,000 pages through Mistral's own API, $2 per 1,000 pages through the Batch API (a 50% discount, asynchronous), and $5 per 1,000 pages for Document AI, per the launch announcement from Mistral AI and its OCR model cards. Microsoft Foundry and Amazon SageMaker bill through their own marketplaces, and Mistral does not publish those rates on its pricing page. Self-hosted enterprise pricing is available only on request from Mistral sales.

Key Features

• Bounding Boxes: Returns a paragraph-level bounding box for every block in reading order when include_blocks is true, which lets RAG pipelines highlight the exact passage a citation came from.
• Typed Block Classification: Labels each block with a structural type, from title, table and equation to header, footer and signature, so downstream code can route tables and signatures separately.
• Per-Word Confidence Scores: Returns confidence at page, block or word granularity through confidence_scores_granularity, so weak extractions can go to human review before they reach an index.
• 170-Language Support: Covers 170 languages across 10 language groups, including low-resource scripts that many competing OCR services support incompletely or not at all.
• Self-Hosted Single-Container Deployment: Deploys as one container inside your own infrastructure for enterprise customers under an agreement with Mistral; no public image or weights download exists.
• Document AI Add-On: Routes OCR 4 output through mistral-small-2603 to produce schema-conformant JSON from any document type, eliminating the need for custom post-processing prompt engineering.

Pros

• Independent annotators preferred OCR 4 over every rival system in Mistral's blind test of 600+ documents, with win rates averaging 72%, per the June 2026 announcement.
• Self-hosted single-container deployment lets regulated industries keep documents inside their own infrastructure, something cloud-only services like Google Document AI and AWS Textract cannot match.
• Broad multilingual coverage extends to low-resource and non-Latin scripts that competing OCR services handle inconsistently.

Cons

• The benchmark evidence is vendor-run: Mistral chose the test documents and notes that OlmOCRBench and OmniDocBench have known scoring limitations.
• Structured output (bounding boxes, block types, confidence scores) requires setting the include_blocks flag explicitly; it is not returned by default, which catches integrators by surprise.
• Self-hosted enterprise deployment requires a commercial sales agreement rather than a self-serve download, adding procurement time for regulated buyers.

Benchmarks

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Mistral OCR 4 Official Page