Existing OCR 3 pipelines
Keep a stable, lower-priced production model while validating quality, output changes, and cost before migrating to OCR 4.1.
Independent tool overview
Mistral OCR 3 at a glance
Mistral OCR 3 is a managed document-extraction model that converts PDFs and images into structured Markdown, tables, embedded images, hyperlinks, and optional schema-constrained annotations for downstream search, analytics, and automation.
Visit the official Mistral OCR 3 site ↗
- Fixed model ID
- mistral-ocr-2512
- Current position
- Supported previous generation
- Recommended successor
- Mistral OCR 4.1
- Standard OCR price
- $2 per 1,000 pages
- Annotated page price
- $3 per 1,000 pages
- Maximum uploaded file
- 512 MB
- Managed data region
- EU data centers by default
- Last reviewed
- August 29, 2026
Overview
What Mistral OCR 3 is
OCR 3 was released in December 2025 for forms, handwriting, low-quality scans, technical documents, and complex tables. Developers can call the OCR API with the fixed model ID mistral-ocr-2512, while nontechnical users can test documents in Mistral Studio’s Document AI interface.
The model preserves reading order and document hierarchy, can separate headers and footers, reconstruct tables as Markdown or HTML, and extract interleaved text and images. Structured annotation requests add a user-defined JSON schema for document-level or image-level fields.
OCR 3 remains supported for existing integrations and production workloads, but it is no longer Mistral’s newest OCR model. The mistral-ocr-latest alias now points to OCR 4.1, which adds paragraph-level bounding boxes, structural block labels, and confidence scores. New projects should benchmark OCR 4.1; existing OCR 3 pipelines should pin mistral-ocr-2512 instead of relying on the moving alias.
Use cases
Who Mistral OCR 3 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Forms and operational documents
Extract labels, boxes, printed values, handwriting, receipts, invoices, compliance forms, and other mixed-layout content.
Document-to-knowledge workflows
Turn reports, manuals, archives, and scientific or technical PDFs into Markdown and images for search, RAG, or agent context.
Schema-based extraction
Request typed JSON fields for invoices, applications, receipts, or document classifications after defining and validating a schema.
Capabilities
Core Mistral OCR 3 features
Layout-aware text extraction
Returns page-level Markdown that preserves headings, paragraphs, lists, equations, and reading order more effectively than plain text OCR.
Complex table reconstruction
Produces inline or separate Markdown tables and HTML tables with merged-cell structures using colspan and rowspan.
Handwriting and mixed forms
Targets cursive writing, handwritten notes over printed content, form boxes, labels, and dense layouts.
Low-quality scan handling
Is designed to tolerate skew, compression artifacts, low DPI, distortion, and background noise better than earlier OCR generations.
Image extraction
Returns embedded images and their positions alongside surrounding text for multimodal document pipelines.
Headers, footers, and hyperlinks
Can separate repeated headers and footers from main content and return link information instead of flattening everything into the body.
Structured annotations
Supports JSON object or JSON schema output for document- and image-level annotations at the annotated-page rate.
Batch API
Processes asynchronous high-volume jobs outside real-time rate limits, with up to 100,000 requests per batch.
Studio and self-hosting options
Offers a drag-and-drop Document AI experience in Mistral Studio and a sales-led self-hosting option for stricter data requirements.
Process
How the Mistral OCR 3 workflow works
Step 1
Choose the version deliberately
Use mistral-ocr-2512 for a stable OCR 3 integration; benchmark OCR 4.1 for new work and never assume mistral-ocr-latest still means OCR 3.
Step 2
Build a representative test set
Include every document type, language, scan quality, table pattern, handwriting style, and failure case that matters in production.
Step 3
Prepare the document
Provide a supported uploaded file, public URL, or Base64 input and split very large documents into traceable chunks when needed.
Step 4
Configure the output
Choose Markdown or HTML table handling, header and footer extraction, image inclusion, page selection, and any annotation schema.
Step 5
Run synchronous or batch OCR
Use real-time requests for interactive jobs and the Batch API for large asynchronous backfills with retry-safe identifiers.
Step 6
Validate every response
Check page counts, missing sections, field types, totals, table shape, reading order, links, image references, and parse errors.
Step 7
Route uncertain documents
Send high-risk, unreadable, incomplete, or contradictory results to a person or a second extraction method before taking action.
Step 8
Measure and monitor
Track field-level accuracy, document failure rate, latency, cost, model version, drift, retention, and downstream corrections.
Cost
Mistral OCR 3 pricing and free plan
OCR 3 is priced by processed page: $2 per 1,000 standard OCR pages or $3 per 1,000 annotated pages. Its launch documentation states that Batch API processing reduced standard OCR to $1 per 1,000 pages. Mistral’s current Free plan includes $10 per month in API credits, subject to the organization’s displayed limits. OCR 4.1 is newer and more expensive at $4 per 1,000 OCR pages or $5 per 1,000 Document AI pages.
Free API credits
$10/month included
Limited Mistral API usage for testing under the Free organization plan.
- Can cover approximately 5,000 OCR 3 pages at the standard rate before other API usage
- Organization rate and document limits apply
- Requires a Mistral Studio organization and API key
- Included credits and limits can change
OCR 3 standard
$2 / 1,000 pages
Text, layout, table, hyperlink, and embedded-image extraction with mistral-ocr-2512.
- Equivalent to $0.002 per page
- Launch documentation lists $1 per 1,000 pages through Batch API
- Batch work is asynchronous
- Actual organization throughput depends on account limits
OCR 3 annotations
$3 / 1,000 pages
OCR plus schema-driven structured document or image annotations.
- Equivalent to $0.003 per annotated page
- Schema design and validation are the customer’s responsibility
- A valid JSON response can still contain incorrect field values
- Downstream model or storage costs are separate
OCR 4.1
$4 OCR or $5 Document AI / 1,000 pages
Current successor with native blocks, structural labels, bounding boxes, and confidence scores.
- mistral-ocr-latest currently points to OCR 4.1
- Batch OCR is currently listed at $0.40 per 1,000 pages
- New projects should compare the higher standard price with the added structure and review controls
- Public Preview status is listed for OCR 4.1
Pricing checked . Check current pricing at the source ↗
Assessment
Mistral OCR 3 strengths and limitations
Where it stands out
- Very low per-page price for managed document extraction
- Handles text, tables, hyperlinks, equations, handwriting, forms, and embedded images in one response
- HTML table output can preserve merged cells and hierarchical table structure
- Fixed model ID allows existing integrations to avoid unplanned upgrades
- Batch processing supports high-volume backfills without consuming real-time rate limits
- Structured annotations reduce the amount of custom parsing needed after OCR
- Self-hosting is available through an enterprise arrangement for organizations that cannot use the managed service
What to consider
- OCR 3 is no longer the newest model; OCR 4.1 has better structural extraction and confidence metadata for new production designs.
- The mistral-ocr-latest alias moves over time and now resolves to OCR 4.1, so it can change output and cost if used without version pinning.
- OCR 3 does not provide OCR 4’s native paragraph block array, structural block labels, or OCR 4.1’s block- and word-level confidence scores.
- OCR can omit characters, transpose digits, merge columns, misread handwriting, or place content in the wrong reading order.
- HTML table structure can look plausible while cells, totals, headers, or row associations are wrong.
- Schema-constrained JSON guarantees shape more reliably than factual correctness; every critical field still needs validation.
- The published 74% win rate over OCR 2 is a vendor-reported internal benchmark and does not predict accuracy on a specific document set.
- Uploaded files can be as large as 512 MB and are retained for 30 days unless deleted earlier, which requires deliberate lifecycle controls for sensitive data.
- Batch results are available for download for 24 hours, so production systems need timely retrieval, idempotency, and failure recovery.
- The managed API uses EU data centers by default, but regional availability, subprocessors, legal basis, and cross-border requirements still need contract review.
- Self-hosting is sales-led rather than an openly downloadable OCR 3 weight release.
- OCR extraction is only one layer; document classification, fraud detection, business-rule validation, review queues, indexing, and long-term storage remain separate.
Compare
Mistral OCR 3 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Business Operations
Mistral OCR 4.1
Mistral OCR 4.1 is the direct successor and adds paragraph blocks, structural labels, bounding boxes, and confidence scores.
Explore Mistral OCR 4.1 →Business Operations
DeepSeek OCR 2
DeepSeek OCR 2 is an open-source alternative for teams willing to operate their own model and evaluate its document-tokenization approach.
Explore DeepSeek OCR 2 →Business Operations
GLM-OCR
GLM-OCR is another document-understanding option for teams comparing multilingual extraction quality, deployment paths, and pricing.
Explore GLM-OCR →Questions
Mistral OCR 3 FAQs
What is Mistral OCR 3?
Mistral OCR 3 is a managed document-extraction model that converts PDFs and images into Markdown, tables, hyperlinks, embedded images, and optional schema-structured annotations.
Is Mistral OCR 3 still available?
Yes. Mistral lists OCR 3 as active and supported for existing integrations and production workloads, although OCR 4.1 is now the recommended newer model.
How much does Mistral OCR 3 cost?
Standard OCR costs $2 per 1,000 pages and annotated extraction costs $3 per 1,000 pages. The OCR 3 launch page lists standard Batch API processing at $1 per 1,000 pages.
What model ID should I use for OCR 3?
Use mistral-ocr-2512 to pin OCR 3. The mistral-ocr-latest alias now points to OCR 4.1 and should not be used when exact OCR 3 behavior is required.
What is the difference between OCR 3 and OCR 4.1?
OCR 4.1 adds native paragraph-level bounding boxes, structural block labels, and page-, block-, or word-level confidence scores. OCR 3 is cheaper at its standard rate and remains supported.
Can OCR 3 extract tables?
Yes. It can return tables as Markdown or HTML, including merged-cell structures. Complex outputs still need checks for missing rows, wrong cell associations, and incorrect totals.
Can OCR 3 read handwriting?
It is specifically designed to improve cursive, handwritten fields, and notes layered over printed forms, but handwriting accuracy varies and should be measured on real samples.
Can OCR 3 return structured JSON?
Yes. Annotations can use JSON object or JSON schema output. The schema constrains format, not the truth of extracted values.
What document formats are supported?
Mistral’s current OCR platform documents PDF, PNG, JPG/JPEG, TIFF, BMP, GIF, and WEBP uploads, with a 512 MB maximum uploaded file size.
Can Mistral OCR 3 be self-hosted?
Mistral offers a self-hosting option through an enterprise engagement for stricter privacy or classified-data requirements. OCR 3 is not presented as an open-weight public download.
Bottom line
Our Mistral OCR 3 verdict
Mistral OCR 3 remains attractive for established pipelines that value its low $2-per-1,000-page price and stable 2512 model ID. It should now be treated as a supported compatibility choice, not the automatic starting point. New projects should compare OCR 4.1’s higher standard price with its block structure, bounding boxes, and confidence scores; regulated or financial workflows should choose based on field-level accuracy and review requirements, not the headline benchmark.
Visit Mistral OCR 3 website ↗Similar Business Operations Tools
View all Business Operations AI tools →
Microsoft Copilot
Microsoft Copilot: Is your ai companion across microsoft 365 for writing, summarizing, and automating work.
Claude Marketplace
Claude Marketplace - Anthropic's enterprise hub for buying Claude-powered partner tools
Shopify SimGym
Shopify SimGym - Simulate buyer behavior with AI shoppers and gain the confidence of high traffic brands before launch.
Voxtral TTS
Voxtral TTS - Mistral's voice cloning model for building multilingual AI speech agents
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.