Mistral OCR 4

Mistral OCR 4 is Mistral AI's document-understanding model that turns PDFs, Office files and scans into structured, citation-ready output: Markdown plus bounding boxes, typed blocks and confidence scores. It is built for teams feeding RAG, enterprise-search and agent pipelines who need accurate extraction in 170 languages at $4 per 1,000 pages, or fully on their own infrastructure.

Mistral OCR 4 is the document-intelligence model Mistral AI released on 23 June 2026, and it changes what the company's OCR line returns: rather than a flat text dump, it produces a structured representation of the whole document, with Markdown text plus bounding boxes, typed blocks (titles, tables, equations, signatures and other structural labels), tables rendered as Markdown or HTML, extracted images, headers, footers and hyperlinks, and confidence scores at page, block or word level. That structure is aimed at RAG, agentic and enterprise-search pipelines that need to cite exactly where on a page an answer came from. Mistral says it reads 170 languages across 10 language groups and accepts PDF, Word, PowerPoint and OpenDocument files as well as PNG, JPEG and AVIF images; a 4.1 point release, dated 16 July 2026 in Mistral's model documentation, added native paragraph-level bounding boxes and block-level confidence. On Mistral's own numbers it scores 85.20 on OlmOCRBench and 93.07 on OmniDocBench and won about 72% of blind head-to-head comparisons judged by independent annotators across more than 600 documents in over 12 languages, although Mistral itself flagged ground-truth errors in the reference sets and VentureBeat noted it sat third on the public OlmOCRBench leaderboard behind open models such as Chandra OCR 2. Pricing is $4 per 1,000 pages, $2 through the batch API and $5 with Document AI annotations. It is served through Mistral's API and Studio, Amazon SageMaker and Microsoft Foundry, with Snowflake Parse Document announced as coming, and enterprise customers can run it as a single self-hosted container, the option regulated buyers comparing it with Amazon Textract, Google Document AI or Azure Document Intelligence will weigh most.

Category

AI Models & APIs

Pricing

Usage-based, Contact for pricing

Target Market

CTOs, Enterprise Developers, Data Scientists, Data Engineers

Deployment

API-based, Self-hosted, Multi-cloud

Founded

2023

Headquarters

Paris, France

Team Size

500+

• ✓
  Structured block output

  Returns typed blocks such as titles, tables, equations and signatures with bounding boxes, so downstream systems can cite the exact region of a page.

• ✓
  Confidence scores

  Emits page-, block- and word-level confidence scores, letting pipelines route low-confidence pages to human review instead of trusting every extraction.

• ✓
  170-language coverage

  Reads 170 languages across 10 language groups, with Mistral claiming gains on rare and low-resource languages that older OCR engines handle poorly.

• ✓
  Self-hosted container

  Enterprise customers can run the model as a single container on their own infrastructure, keeping sensitive documents inside regulated environments.

• ✓
  Batch API pricing

  Halves the price to $2 per 1,000 pages for asynchronous jobs, which matters when backfilling very large document archives.

• ✓
  Document AI annotations

  For $5 per 1,000 pages, adds custom JSON-schema structuring, image annotation and prompt-driven processing on top of the raw OCR output.

• ✓
  Multi-format input

  Accepts PDF, DOCX, PPTX, OpenDocument and PNG, JPEG or AVIF images by URL, base64 or upload, avoiding a separate conversion step.

• •
  Citation-ready RAG ingestion

  Parsing contracts, filings and manuals into typed blocks with coordinates so a retrieval system can show users the exact passage and page it quoted.

• •
  Financial document QA

  Extracting tables and figures from financial reports at scale; Rogo reported equivalent accuracy at about 8x lower cost and 17x lower latency than agentic parsers, per Mistral's announcement.

• •
  IP docketing workflows

  Processing patent and legal correspondence in docketing workflows; Anaqua reported roughly 4x faster per-page processing than its incumbent provider, per Mistral's announcement.

• •
  On-premises regulated processing

  Digitising sensitive records inside a bank, insurer or public agency by running the self-hosted container, so documents never leave the organisation's network.

• •
  Multilingual archive digitisation

  Batch-converting scanned archives in dozens of languages into searchable Markdown at $2 per 1,000 pages through the batch API.

Metered per page, not per token: $4 per 1,000 pages on the standard API, $2 via batch and $5 with Document AI annotations, all published. Self-hosted container pricing is not published and is negotiated as an enterprise deal. Mistral's Free plan includes $10 a month of general API credits.

This page was written from 12 sources, 10 on domains other than mistral.ai.

1. 1.mistral.ai — ocr 4vendor
2. 2.docs.mistral.ai — basic ocr
3. 3.docs.mistral.ai — ocr 4 1
4. 4.mistral.ai — pricingvendor
5. 5.help.mistral.ai — 347638 do you have soc 2 or iso 27001 certification
6. 6.venturebeat.com — mistral launches ocr 4 turning document extraction into a fu
7. 7.marktechpost.com — mistral ocr 4
8. 8.tech-insider.org — mistral ocr 4 launch 2026
9. 9.news.ycombinator.com — item
10. 10.news.ycombinator.com — item
11. 11.techcrunch.com — mistral raises e3b as sovereign ai becomes big business
12. 12.en.wikipedia.org — Mistral AI