Mistral OCR 4: enterprise document extraction with bounding boxes, confidence scores, and 170 languages

What's new

• Bounding boxes and block classification: returns block types (titles, tables, equations, signatures) with bounding boxes
• Per-word confidence scores: inline confidence scores for every word
• 170 languages: including less common languages
• Self-hosting: single Docker container for on-premise deployment — sensitive data stays within customer infrastructure
• Distribution: Mistral API, Amazon SageMaker, Microsoft Foundry
• Pricing: $4 per 1,000 pages ($2 with Batch API discount)

Why it matters

OCR 4 addresses a key barrier for regulated enterprises (healthcare, finance, legal) — none of the leading US AI providers (OpenAI, Anthropic, Google) can offer a fully self-hosted document AI with comparable performance. Mistral targets the segment where European regulation and GDPR rule out cloud solutions.

How to try it

API access at mistral.ai/news/ocr-4. Batch API available for reduced pricing. Self-hosting documentation at docs.mistral.ai.