AI

Mistral OCR 4 released with structured-document output

2026-06-23 digest confirmed

  • Category: AI
  • Status: confirmed
  • Sources: Mistral, discussion
  • Summary: Mistral released OCR 4 on 2026-06-23 (model id mistral-ocr-latest). It moves past plain text extraction to return a structured document representation: text with bounding boxes, block classification (titles, tables, equations, signatures), and per-element confidence scores, across 170 languages and PDF/DOC/PPT/OpenDocument inputs. Mistral reports OlmOCRBench 85.20 (highest among the models it tested), OmniDocBench 93.07, and a human-preference evaluation with average 72% win rates over competing systems, while noting ground-truth and formatting artifacts limit such benchmarks. API pricing is $4 per 1,000 pages ($2 batch, $5 Document AI); available via Mistral Studio, Amazon SageMaker, Microsoft Foundry, and self-hosting for enterprise.
  • Why it matters: Structured OCR output with localized elements and confidence scores feeds RAG and agent document pipelines directly, and the per-page pricing sets a concrete cost point for document-ingestion workloads.
  • Follow-up: Watch for independent OlmOCRBench/OmniDocBench reproduction and self-hosted throughput figures.

Send feedback on this story