Mistral AI just dropped a pricing bomb on the Document AI market. Their new Mistral OCR 3 (mistral-ocr-2512) extracts structured data from complex documents at $2 per 1,000 pages—that's a 97% cost reduction compared to AWS Textract's forms and tables tier.

What's New

Released on December 18, 2025, OCR 3 is Mistral's answer to enterprise document processing pain points. The model handles the documents that typically break OCR systems: handwritten notes, dense multi-column forms, skewed scans, and complex tables with merged cells.

According to Mistral's announcement, OCR 3 achieves a 74% win rate over its predecessor (OCR 2) on forms, scanned documents, complex tables, and handwriting. The output? Clean markdown with full HTML table reconstruction, preserving colspan and rowspan attributes—critical for downstream RAG pipelines and data extraction workflows.

Key Features

  • Markdown + HTML Output: Structured extraction that maintains document layout fidelity, including nested tables and embedded images
  • Complex Document Handling: Purpose-built for forms with complex layouts, scanned documents with noise, and cursive handwriting
  • Batch API: High-volume processing at 50% discount ($1/1,000 pages) for bulk workloads
  • JSON Mode: Native support for structured JSON output via response_format parameter
  • Backward Compatibility: Drop-in replacement for OCR 2 workflows

For Developers

Here's what matters for your implementation:

The model is accessible via the standard Mistral API using the mistral-ocr-2512 model identifier. If you're processing archived PDFs or running digitization projects, the Batch API is where you'll see real savings—cutting costs to $1 per 1,000 pages for asynchronous processing.

from mistralai import Mistral

client = Mistral(api_key="your_api_key")

response = client.ocr.process(
    model="mistral-ocr-2512",
    document={"url": "https://example.com/invoice.pdf"},
    response_format="markdown"
)

print(response.content)  # Structured markdown output

For RAG implementations, the markdown output slots directly into chunking pipelines. The HTML table preservation means you won't lose critical tabular relationships during vector embedding—a common failure point with generic OCR solutions.

Comparison: How It Stacks Up

Let's cut through the marketing and look at real numbers. The Document AI market is projected to reach $27.62 billion by 2030, growing at 13.5% CAGR. Here's how Mistral positions against the incumbents:

ServicePrice per 1,000 PagesBatch PricingKey Differentiator
Mistral OCR 3$2$1Markdown/HTML output, aggressive pricing
AWS Textract$65 (forms/tables)N/AEnterprise SLAs, AWS ecosystem integration
Google Document AI$30-45Volume discountsPre-trained processors, GCP native
Azure (Mistral hosted)$1-3N/AAzure compliance, hybrid deployment

The 97% cost reduction versus AWS Textract is real, but context matters. AWS and Google offer enterprise-grade SLAs, SOC2/HIPAA compliance documentation, and native cloud integrations that Mistral hasn't publicly matched yet. If you're in regulated industries, factor in compliance requirements before switching.

Getting Started

  1. Get API Access: Sign up at console.mistral.ai and grab your API key
  2. Test in Playground: Use the Document AI Playground in Mistral AI Studio to evaluate quality on your document types
  3. Review Documentation: Check the OCR 3 docs for endpoint specs and response formats
  4. Start with Standard API: Validate accuracy before committing to batch workflows
  5. Scale with Batch: Move high-volume processing to the Batch API for 50% savings

Verdict

Mistral OCR 3 is a legitimate disruptor for cost-conscious teams processing documents at scale. The $2/1,000 pages pricing (or $1 via Batch API) makes previously expensive digitization projects economically viable—especially for startups and mid-market companies building RAG systems or automating document workflows.

The 74% improvement over OCR 2 is promising, but these are internal benchmarks. Independent validation is still pending since the December 18 launch. Our recommendation: run pilot tests on your specific document types before committing production workloads.

Best for: Teams building RAG pipelines, document digitization projects, invoice/form processing at scale, and anyone currently paying enterprise rates for basic OCR.

Hold off if: You need documented compliance certifications (SOC2, HIPAA), guaranteed SLAs, or deep cloud-native integrations with AWS/GCP services.

At this price point, the barrier to experimentation is essentially zero. Spin up a test, throw your worst documents at it, and see if the output quality justifies migrating from your current solution.