Enterprise search and RAG
Create citation-ready chunks from mixed-layout documents while preserving page, region, and structural context.
Independent tool overview
Mistral OCR 4.1 at a glance
Mistral OCR 4.1 is a document-extraction model that returns text plus layout-aware blocks, bounding boxes, structural labels, tables, images, and confidence scores for search, RAG, automation, and human review.
Visit the official Mistral OCR 4.1 site ↗
- Current version
- OCR 4.1
- API model
- mistral-ocr-latest
- Standard API price
- $4 per 1,000 pages
- Document AI price
- $5 per 1,000 annotated pages
- License
- Premier
- Deployment
- Mistral API, partner clouds, or enterprise self-hosting
Overview
What Mistral OCR 4.1 is
Mistral OCR 4 introduced a more structured form of document parsing in June 2026. OCR 4.1 followed in July and is now the current model behind the aliases mistral-ocr-latest and mistral-ocr-4. The update adds block-level confidence scoring to the paragraph bounding boxes and structural block labels introduced with 4.0.
The service converts PDFs, office documents, and images into page-level Markdown and structured metadata. When block extraction is enabled, each region is returned in reading order with coordinates and a type such as title, text, list, table, image, equation, caption, code, reference, header, footer, aside, or signature.
Developers can call the OCR endpoint directly for custom ingestion pipelines or use Mistral Document AI for application-level structured extraction. OCR 4 also feeds search, retrieval, agent, invoice, compliance, redaction, and citation workflows because downstream systems can use both what a region says and where it appears.
OCR output is evidence to verify, not a source of truth. Confidence scores help route uncertain material, but they do not prove correctness. Production systems should preserve the original document, link every extracted value to its page and bounding box, validate business-critical fields, and send low-confidence or high-risk cases to a human.
Use cases
Who Mistral OCR 4.1 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Document automation
Extract invoices, forms, contracts, reports, presentations, and other records into typed downstream workflows.
Multilingual archives
Process document collections spanning many scripts and language groups through one OCR interface.
Regulated review workflows
Use coordinates and confidence scores to highlight source regions, route uncertainty, support redaction, and keep humans in the loop.
Data-sovereign deployments
Discuss single-container self-managed deployment when documents cannot leave the organization's controlled environment.
Capabilities
Core Mistral OCR 4.1 features
Layout-preserving text extraction
Returns page Markdown while preserving hierarchy, tables, images, hyperlinks, and other document structure.
Paragraph bounding boxes
Each extracted block can include page coordinates so applications can highlight the exact source region or build grounded citations.
Structural block labels
Classifies regions as text, title, list, table, image, equation, caption, code, references, aside, header, footer, or signature.
Confidence scores
OCR 4.1 supports page-, block-, or word-level confidence output for risk-based review and quality monitoring.
Table and image extraction
Tables can be returned inline or separately in Markdown or HTML, while extracted images can be mapped back to placeholders.
Header and footer separation
Optional parameters move recurring headers and footers into their own response fields instead of mixing them with main content.
Structured annotations
A prompt and JSON response format can extract document-level fields into a caller-defined structure.
Page selection
Process all pages or specify page indexes and ranges for targeted extraction and lower cost.
Batch processing
Asynchronous batch inference supports high-volume document jobs without consuming real-time rate limits.
Multilingual document understanding
Mistral says the OCR 4 family supports 170 languages across 10 language groups, including specialized and lower-resource languages.
Flexible deployment
Available through Mistral Studio and API, Amazon SageMaker, Microsoft Foundry, and enterprise self-managed deployment.
Process
How the Mistral OCR 4.1 workflow works
Step 1
Define the output contract
List required text, tables, fields, coordinates, confidence thresholds, and failure states before choosing basic OCR or Document AI annotations.
Step 2
Build a representative evaluation set
Sample real scans, languages, handwriting, forms, multi-column pages, tables, equations, stamps, signatures, and low-quality inputs from the target workload.
Step 3
Choose the deployment path
Use the stateless OCR API for custom pipelines, Document AI for structured application workflows, Batch for volume, or discuss self-hosting for sovereignty.
Step 4
Request source-grounded output
Enable blocks, confidence scores, table formatting, and header or footer separation, then retain page indexes and bounding boxes with every extracted value.
Step 5
Validate and route uncertainty
Apply schema, type, checksum, range, and cross-field checks; send low-confidence or high-impact fields to a human with the original page visible.
Step 6
Monitor drift and cost
Track accuracy by document type and language, review false confidence, compare pinned and latest model aliases, and reconcile processed-page usage.
Cost
Mistral OCR 4.1 pricing and free plan
Mistral prices OCR 4.1 per page rather than per token: $4 per 1,000 pages for the standard OCR API and $5 per 1,000 annotated pages for Document AI. Batch processing is advertised at half the standard API price. Regional, priority, enterprise, partner-cloud, and self-hosted arrangements can change the effective cost.
OCR API
$4 per 1,000 pages
Standard synchronous OCR 4.1 processing through Mistral's API.
- Text, tables, images, layout metadata, and optional blocks
- Optional page-, block-, or word-level confidence scores
- Use mistral-ocr-latest for the current model
- Usage is billed by processed page
Batch OCR
$2 per 1,000 pages
Asynchronous high-volume processing at the advertised 50% Batch discount.
- Does not consume real-time rate limits
- Maximum batch file size is 512 MB
- Up to 100,000 requests per batch
- Completion time depends on queue depth and request complexity
- Batch results remain available for download for 24 hours
Document AI
$5 per 1,000 annotated pages
Application-level structured extraction using OCR plus caller-defined annotations.
- Extract fields into a supplied JSON format or schema
- Better fit for no-code or business-document workflows
- Validate every generated field against the source document
Enterprise and self-hosted
Contact sales
Custom deployment, support, capacity, privacy, and residency arrangements.
- Single-container self-managed deployment is available to enterprise customers
- Regional inference adds 10% where supported
- Enterprise APIs with SLAs, higher limits, regional controls, and premium support may carry a 75% premium on select APIs
- Partner-cloud pricing and terms may differ from Mistral's direct API
Pricing checked . Check current pricing at the source ↗
Assessment
Mistral OCR 4.1 strengths and limitations
Where it stands out
- Returns document structure, locations, and confidence information rather than only plain text.
- Supports source highlighting, citations, redaction, and review workflows through page-aware bounding boxes.
- Typed blocks reduce the amount of downstream layout reconstruction required for search and automation.
- Per-page API pricing is straightforward and low for high-volume text extraction.
- Batch, regional, partner-cloud, Document AI, and self-hosted options cover several operational models.
- Mistral openly describes benchmark artifacts and recommends evaluating the model on a customer's own documents.
What to consider
- OCR 4.0 has already been superseded by OCR 4.1; applications using a latest or family alias should expect behavior to change and keep regression tests.
- The model is Premier rather than an openly downloadable model, and self-managed deployment requires an enterprise agreement.
- Confidence scores are model estimates, not calibrated guarantees; a confident extraction can still be wrong.
- Structured annotations are generated outputs and can omit fields, normalize values incorrectly, or return plausible content not supported by the page.
- Vendor benchmarks are directional, competitor results include Mistral's internal reproductions, and aggregate scores can hide failures on a specific document type or language.
- Complex tables, equations, multi-column reading order, low-resolution scans, handwriting, stamps, signatures, and unusual layouts require workload-specific testing.
- Uploaded files can be retained for 30 days unless deleted earlier; zero data retention is a separate paid-plan control that must be requested and does not cover file uploads or batch files.
- The stateless OCR endpoint can support zero data retention, but Document AI, files, libraries, and batch workflows have different storage behavior.
- Page-based pricing excludes engineering, validation, exception review, storage, indexing, downstream model calls, regional premiums, and partner-platform charges.
- Extracted text and fields can carry personal, confidential, copyrighted, or regulated information, so access, retention, redaction, audit, and deletion controls remain necessary.
Compare
Mistral OCR 4.1 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Business Operations
DeepSeek OCR 2
Choose DeepSeek OCR 2 when an open model and token-efficient document parsing are more important than Mistral's managed Document AI stack.
Explore DeepSeek OCR 2 →Business Operations
GLM-OCR
Choose GLM-OCR when Z.ai's API or deployment ecosystem better matches the surrounding application.
Explore GLM-OCR →Business Operations
HunyuanOCR
Choose HunyuanOCR when evaluating an open-source visual document-understanding model, especially for Chinese-language workloads.
Explore HunyuanOCR →Questions
Mistral OCR 4.1 FAQs
What is Mistral OCR 4.1?
Mistral OCR 4.1 is the current model in the OCR 4 family. It extracts page text and structure and can return paragraph bounding boxes, block types, tables, images, and page-, block-, or word-level confidence scores.
What happened to Mistral OCR 4.0?
OCR 4.0 launched on June 23, 2026. OCR 4.1 followed on July 16 with block-level confidence scores. Mistral's mistral-ocr-latest and mistral-ocr-4 aliases now point to 4.1.
How much does Mistral OCR 4.1 cost?
The standard OCR API costs $4 per 1,000 pages. Mistral advertises Batch at half price, or $2 per 1,000 pages. Document AI structured annotations cost $5 per 1,000 pages.
What does Mistral OCR return?
The response can include page Markdown, images, tables, hyperlinks, headers, footers, page dimensions, confidence scores, usage details, and a reading-ordered list of typed blocks with bounding boxes.
Which document formats are supported?
Mistral documents support for PDFs and common image formats through the OCR flow, with office-document support available through document URLs and processing paths. Current file-upload documentation lists PDF, PNG, JPG, JPEG, TIFF, BMP, GIF, and WEBP.
Does Mistral OCR support tables and equations?
Yes. Tables can be extracted inline or separately as Markdown or HTML, and OCR 4 block labels include table and equation regions. Accuracy still needs testing on the exact table and mathematical layouts in production.
What are OCR 4 block labels?
Available types include text, title, list, table, image, equation, caption, code, references, aside text, header, footer, and signature.
Can Mistral OCR 4.1 be self-hosted?
Mistral says the compact OCR 4 family can run in a single container and offers self-managed deployment to enterprise customers. Pricing and licensing require a sales agreement.
Does Mistral offer zero data retention for OCR?
The stateless /v1/ocr endpoint is eligible for zero data retention on paid plans after approval. ZDR is separate from training opt-out and does not cover uploaded files, batch files, libraries, agents, or other stateful products.
Is Mistral OCR accurate enough for invoices or compliance documents?
It can be a strong extraction component, but no OCR model should directly approve payments, filings, legal obligations, or compliance decisions. Use validation rules, source-linked coordinates, confidence thresholds, and human review for consequential fields.
Bottom line
Our Mistral OCR 4.1 verdict
Mistral OCR 4.1 is a strong managed option for teams that need more than plain transcription. Its combination of Markdown, tables, typed regions, bounding boxes, and granular confidence scores is well suited to citation-ready search, RAG, redaction, and document-automation pipelines. The low page price is attractive, but the real production decision should be based on accuracy by document class, exception-review cost, retention requirements, deployment constraints, and whether the organization needs an open model or a contracted self-hosted option.
Visit Mistral OCR 4.1 website ↗Similar Business Operations Tools
View all Business Operations AI tools →
Meta Business Agent
Meta Enterprise Agent - AI agent for customer sales and support across Meta's apps
Claude Tag
Claude Tag - Tag @Claude as a shared Slack teammate to delegate tasks
Command A+
Command A+ - Cohere's new open-source agentic model combining multimodal, agentic, and multilingual capabilities previously split across separate Command models.
Mercury 2
🗣️ Mercury 2 - Inception's diffusion reasoning model built for realtime voice 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.