Mistral OCR 4 points at a better test for document AI than “can it read the page?” The production question is whether the system can explain what it read, where it found it, and how confident it is before a downstream workflow acts on the result.
The release adds bounding boxes, typed block classification, inline confidence scores, and Markdown-structured extraction. Mistral says the model supports 170 languages across 10 language groups, accepts common enterprise formats such as PDF, DOC, PPT, and OpenDocument, and can run in a single container for self-hosted deployments. Pricing is listed at $4 per 1,000 pages through the API and $2 per 1,000 pages through batch.
Grey Haven’s read: OCR is becoming ingestion infrastructure for regulated and operational workflows. In legal docketing, healthcare records, insurance claims, logistics documents, procurement, and finance, the hard part is not turning a PDF into text. The hard part is preserving enough structure and confidence for review, citation, redaction, and exception handling.
That makes the self-hosting detail important. Some customers will not send sensitive document flows to a black-box service if residency, auditability, or volume economics are non-negotiable. A compact deployment model gives operators another path between brittle legacy OCR and fully managed document AI.
The operator watch item is verification. Teams should test OCR systems on their ugliest documents, not vendor samples: scans, stamps, signatures, multilingual pages, weird tables, rotated pages, and low-quality faxes. Confidence scores and bounding boxes only matter if they help route uncertain cases to humans instead of laundering bad extraction into a clean-looking answer.
Source: Mistral AI, “Mistral OCR 4: SOTA OCR for Document Intelligence,” June 23, 2026.
