Mistral releases OCR 4.1 with native paragraph-level bounding box extraction

What Mistral OCR 4.1 adds

Mistral OCR 4.1, released in public preview on 14 August 2026, extends the company’s document-understanding API with three structural features: native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores. Pricing in preview is €3.50 per 1,000 pages.

The shift is from text extraction to document structure. Earlier OCR pipelines answered “what does this page say?”; 4.1 also answers “where on the page does each piece live, what kind of block is it, and how sure is the model about it?” — which is precisely the metadata an automated consumer of the output needs.

Why bounding boxes matter for agentic workflows

When a human reads an OCR transcript, lost layout is an inconvenience. When an AI agent consumes it, lost layout is lost meaning: an invoice total, the cell of a table, and a footnote all flatten into undifferentiated text. Paragraph-level bounding boxes keep the link between each extracted block and its physical position, so an agent can cite where a figure came from, cross-check a claim against the region of the page that contains it, and hand a human reviewer a highlighted source rather than a wall of text.

That traceability is what separates a document agent you can audit from one you have to trust blindly. Grounding answers in a specific region of a specific page is the document-processing equivalent of citing sources.

What block labels and confidence scores change

Structural block labels tell the consuming agent what each region is — so a downstream pipeline can treat tables, headings, and body text differently instead of parsing one undifferentiated stream. Block-level confidence scores enable selective human review: route only the low-confidence blocks to a person, and let the rest flow straight through. For document-heavy workflows — claims processing, contract review, compliance intake — that single change is often the difference between a pilot and a production deployment, because it caps the review burden at the uncertain fraction of each document.

How it fits a document-heavy agent pipeline

OCR sits at the front of most document agent architectures: files come in, structured text comes out, and everything downstream — retrieval, extraction, decision-making — works from that output. Structured, position-aware, confidence-scored output makes the downstream stages simpler and more testable: retrieval can index blocks instead of arbitrary chunks, extraction can validate against block types, and evaluation can measure accuracy per block rather than per document.

What to weigh before adopting

It is a public preview, so treat the API surface and pricing as subject to change before committing a production workload. The €3.50 per 1,000 pages figure makes cost modelling straightforward — price out your monthly document volume and compare it against your current OCR spend plus the human review time the confidence scores would save.

And evaluate on your own documents before switching: OCR quality varies sharply with scan quality, language, and layout complexity, so a benchmark on someone else’s corpus tells you little. A golden set of your hardest fifty pages, scored per block, will tell you more than any launch post.