Mistral OCR 3 Review: 74% Better Than OCR 2, $2 per 1k Pages (2026)
Released December 17, 2025 · Reviewed June 14, 2026
At a Glance
On December 17, 2025, Mistral AI released Mistral OCR 3 — the third generation of their document intelligence model. The headline number is a 74% overall win rate over Mistral OCR 2 on forms, scanned documents, complex tables, and handwriting. At $2 per 1,000 pages ($1 with Batch API), it enters a competitive market against AWS Textract, Azure Document Intelligence, and AI-native OCR solutions.
Alongside the model, Mistral launched the Document AI Playground in Mistral AI Studio — a drag-and-drop interface for parsing PDFs and images into clean text or structured JSON without writing any code. The model ID is mistral-ocr-2512 and it is backward compatible with Mistral OCR 2.
What's New vs. Mistral OCR 2
74% win rate across real-world document types
Mistral benchmarked OCR 3 against OCR 2 on real business document examples from customers — forms, handwritten content, low-quality scans, and complex tables. OCR 3 wins 74% of those comparisons overall. Mistral also reports outperforming enterprise document processing solutions (likely Adobe, AWS Textract, Azure Document Intelligence) and AI-native OCR products on accuracy. The benchmarks use a fuzzy-match metric against ground truth, which is more forgiving of minor formatting variations than strict character-level matching.
HTML table reconstruction that preserves complex structure
Most OCR solutions flatten tables into plain text, losing column hierarchies, merged cells, and multi-row blocks — forcing downstream systems to reconstruct structure from whitespace. Mistral OCR 3 outputs HTML table tags with colspan and rowspan attributes, preserving the logical structure of the original document. This is particularly valuable for financial tables, compliance reports, and any document where which cell belongs to which header is semantically critical. The HTML is embedded in markdown output, making it compatible with markdown-aware parsers while preserving structural fidelity.
Handwriting and mixed-content recognition
Handwriting is the traditional weak point of OCR systems, particularly for cursive, annotated forms, and handwritten text layered over printed content. Mistral OCR 3 significantly improves handling of cursive writing, mixed-content annotations, and handwritten entries within printed form fields. For organizations digitizing paper archives — medical records, historical documents, compliance filings — this means meaningfully fewer manual corrections. The improvement extends to low-quality scans with compression artifacts, skew, distortion, and low DPI.
Document AI Playground for non-developer access
Alongside the API, Mistral released the Document AI Playground in Mistral AI Studio — a simple drag-and-drop interface for parsing PDFs and images into clean text or structured JSON. This lets business analysts, operations teams, and non-developers process documents without writing code. For developers, it's a quick way to test OCR 3 output quality before committing to API integration. The playground is available in the same Mistral Studio environment that hosts their audio playground and model testing interfaces.
Competitive pricing with meaningful Batch API discount
At $2 per 1,000 pages, Mistral OCR 3 positions itself as affordable relative to enterprise alternatives. AWS Textract charges per page with different rates for different feature tiers; Azure Document Intelligence has similar per-page pricing that can run $1.50–$5 depending on the service tier. The Batch API discount (50%, reducing to $1/1k pages) makes high-volume processing significantly cheaper — invoice processing pipelines or archive digitization projects that handle millions of pages per month will see major cost differences.
Comparison
Accuracy comparisons from Mistral's published benchmarks. Pricing approximate as of December 2025.
| Model | Accuracy | Price | Tables | Handwriting |
|---|---|---|---|---|
| Mistral OCR 3 | 74% win vs OCR 2 | $2/1k pages | HTML (colspan/rowspan) | Yes |
| Mistral OCR 2 | Baseline | $2/1k pages* | Basic | Limited |
| AWS Textract | Strong on forms | $1.50–$15/1k | Yes | Limited |
| Azure Doc Intelligence | Strong on forms | $1.50–$5/1k | Yes | Limited |
| Google Cloud Vision | Good general OCR | $1.50/1k | Basic | Limited |
Source: Mistral AI benchmark publication, December 2025. Enterprise pricing varies; check providers for current rates. *Mistral OCR 2 pricing may differ — check mistral.ai for current rates.
Pricing
Standard API
API$2 / 1,000 pages
Full Mistral OCR 3 API access. Supports PDF and image input. Outputs markdown with embedded HTML tables. Access via Mistral La Plateforme. Test in Document AI Playground.
Batch API
Batch$1 / 1,000 pages
50% discount for asynchronous batch processing. Ideal for high-volume document pipelines — archive digitization, invoice processing, bulk PDF parsing.
Self-hosted
EnterpriseCompute cost only
On-premise deployment available for data-privacy-sensitive use cases (GDPR, classified data, healthcare). Contact Mistral AI for enterprise self-hosting options.
Use Cases
- →Invoice and receipt processing: High-volume extraction of line items, totals, and vendor information from scanned invoices into structured fields for ERP or accounting systems.
- →Compliance document digitization: Government forms, regulatory filings, and compliance documents with mixed handwritten and printed content processed into clean text or JSON.
- →Archive digitization: Historical records, handwritten ledgers, and annotated documents converted to searchable text. Batch API pricing ($1/1k pages) makes large-scale archiving cost-effective.
- →Document-to-knowledge pipelines: Technical reports, scientific papers, and multi-page PDFs parsed into markdown for ingestion into vector databases, RAG systems, or knowledge bases.
- →Healthcare record digitization: Medical forms, handwritten clinical notes, and lab reports processed on-premises via self-hosted deployment for HIPAA compliance.
Frequently Asked Questions
What is Mistral OCR 3?
Mistral OCR 3 is Mistral AI's third-generation document intelligence model, released December 17, 2025. It extracts text and embedded images from PDFs and scanned documents, outputting markdown enriched with HTML-based table reconstruction. The model ID is mistral-ocr-2512. It achieves a 74% win rate over Mistral OCR 2 across real-world document types including forms, handwriting, scanned documents, and complex tables.
How does Mistral OCR 3 pricing work?
Mistral OCR 3 is priced at $2 per 1,000 pages via the standard API. Using the Batch API reduces this to $1 per 1,000 pages (50% discount). Pricing is per page, not per token, which differs from standard LLM API billing. Self-hosting is available for organizations with stringent data privacy requirements — contact Mistral for enterprise arrangements.
What is the Document AI Playground?
The Document AI Playground is a no-code interface in Mistral AI Studio that lets users drag and drop PDFs or images and receive the parsed content as clean text or structured JSON. It uses Mistral OCR 3 under the hood. It's useful for testing OCR quality before committing to API integration, and for non-developer use cases where business teams need to process individual documents without writing code.
How does Mistral OCR 3 handle tables?
Mistral OCR 3 outputs HTML table tags with colspan and rowspan attributes to preserve the structure of complex tables — including headers, merged cells, multi-row blocks, and column hierarchies. This is embedded in the markdown output. Plain text OCR typically loses this structure; downstream systems then have to reconstruct it from whitespace, which is error-prone. The HTML-enriched output is compatible with any markdown-aware parser.
Is Mistral OCR 3 backward compatible with Mistral OCR 2?
Yes. Mistral OCR 3 is fully backward compatible with Mistral OCR 2. Existing integrations using the Mistral OCR API can migrate to OCR 3 without code changes — the model ID (mistral-ocr-2512) is used to request OCR 3 specifically, or you can use the mistral-ocr-latest alias.
Can I self-host Mistral OCR 3?
Self-hosting is available for organizations with stringent data privacy requirements — GDPR-sensitive or classified data scenarios where documents cannot be sent to a third-party API. This is an enterprise arrangement through Mistral AI, not open-weights self-hosting. Contact Mistral for details.
Verdict
Mistral OCR 3 is a genuine upgrade over OCR 2 — a 74% win rate across real customer document types is not a marginal improvement. The HTML table reconstruction differentiates it from commodity OCR: most pipelines that process financial or compliance documents need structure, not just text, and preserving colspan/rowspan in output means less downstream cleanup.
At $2/1k pages ($1 batch), it competes favorably on price against AWS Textract and Azure Document Intelligence for standard document processing workloads. The Document AI Playground lowers the barrier for non-developer adoption. Self-hosting covers the cases where data can't leave your infrastructure.
The main caveat: Mistral's benchmarks are self-published, and independent evaluations are limited at this stage. For mission-critical pipelines, running a parallel evaluation on your actual document corpus before fully migrating is worthwhile.