Mistral ships OCR 4.1 with paragraph-level bounding boxes and block confidence scores

Mistral has put OCR 4.1, the latest version of the OCR engine behind its Document AI stack, into public preview, adding native paragraph-level bounding box extraction and confidence scoring at the block level. The model, identified as mistral-ocr-4-1, became available on Mistral's Premier tier on July 16, 2026.

What's new

Mistral's own documentation describes OCR 4.1 as "our latest OCR service powering our Document AI stack, with native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores." In practice that means three capability additions layered on top of standard text extraction:

• Paragraph-level bounding boxes — geometric coordinates for each paragraph, not just lines or words
• Structural block labels — classification of document regions (e.g., headings, body text, tables)
• Block-level confidence scores — a reliability score attached to each extracted block, letting downstream systems flag low-confidence regions for review

The model is exposed through the existing /v1/ocr endpoint (covering bounding-box extraction, standard OCR, and structured annotations) and /v1/batch for bulk document processing. Pricing is $4 per 1,000 pages for standard OCR, rising to $5 per 1,000 pages when structured annotation output is requested. It is currently in Public Preview rather than general availability.

Context

OCR and document-intelligence tooling has become a competitive front among frontier labs and cloud providers, sitting alongside offerings like AWS Textract and Google Document AI. Mistral has iterated on its OCR line as part of a broader Document AI push aimed at developers building retrieval-augmented generation and data-extraction pipelines that need structured, machine-readable output rather than raw text dumps. The jump to block-level confidence scoring and paragraph-granularity boxes is an incremental but concrete step up from coarser line- or word-level extraction offered by earlier versions.

Why it matters

Most production document pipelines don't just need text out of a PDF or scan — they need to know where that text sits on the page and how much to trust it. Paragraph-level bounding boxes make it easier to reconstruct document layout for downstream rendering or citation, while block-level confidence scores give engineers a mechanical way to route uncertain extractions to human review or a secondary model instead of silently propagating OCR errors into an agent's context window.

For teams building document-heavy agentic workflows — contract review, invoice processing, compliance document ingestion — this kind of structural fidelity is often the difference between an OCR layer that's usable in production and one that needs a custom post-processing layer bolted on top. It's a narrow, infrastructure-level release rather than a headline model launch, but it reflects the steady buildout of the data-ingestion layer that increasingly sits underneath higher-profile agentic products.