French AI company Mistral released OCR 3 this week, the third iteration of its document recognition system. The pricing caught attention: $2 per 1,000 pages, dropping to $1 with batch processing discounts.
Mistral reports a 74% win rate against "competing products" on forms, scanned documents, complex tables, and handwritten content. That figure comes from Mistral's internal benchmarks, using fuzzy-match accuracy against ground truth. Independent verification hasn't emerged yet. IDC's Tim Law offered the expected industry endorsement, calling OCR "foundational for enabling generative AI and agentic AI," though that's the standard analyst framing for any document processing release.
The technical story is more concrete. OCR 3 outputs markdown with HTML table reconstruction (colspan, rowspan tags intact), meaning downstream systems can parse structure without extra cleanup. It handles cursive handwriting, mixed annotations over printed forms, compression artifacts, skew, and low DPI scans. Mistral built a Document AI Playground into their Studio interface for drag-and-drop testing.
This slots into a broader December push from Mistral. The company launched its Mistral 3 model family and Devstral 2 coding tools in recent weeks. Mistral closed a roughly $2 billion Series C in September led by ASML, though that still trails far behind American rivals.
The Bottom Line: Mistral is betting that enterprise AI adoption bottlenecks at document processing, not chatbots, and priced OCR 3 to grab that market early.
QUICK FACTS
- Price: $2 per 1,000 pages ($1 with batch discount)
- Model ID: mistral-ocr-2512
- Win rate: 74% over prior version (company-reported)
- Deployment: API, Document AI Playground, on-premises option
- Backward compatible with OCR 2
