Glonce

Mistral OCR 4.1 ships with bounding boxes, block labels

modelsThe Glonce editorsHacker News🔥70Glonce rating70Impact60Novelty50Relevance88Surprise32Players58Credibility88
Mistral OCR 4.1 ships with bounding boxes, block labels

Mistral AI has released OCR 4.1, the latest version of the OCR service that powers its Document AI stack, into Public Preview as of July 16, 2026. The model's documentation page tags it a "Premier" model in Mistral's catalog and lists three identifiers for calling it: mistral-ocr-4-1, mistral-ocr-4, and mistral-ocr-latest. It costs €3.5 per 1,000 pages of standard output, rising to €4.38 per 1,000 pages, about 25% more, when the output is annotated.

The headline additions in this version are native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores: together they return not just the extracted text but where each element sits on the page, what kind of structural block it belongs to, and how confident the model is in each block's extraction. The documentation page splits these capabilities into four listed features. BBox Extraction, plain OCR, and structured annotations are all reachable through the same /v1/ocr endpoint, while batch processing is exposed separately through a /v1/batch endpoint.

The page does not publish any accuracy, throughput, or latency numbers for OCR 4.1. Its "Speed" and "Performance" rows render only as filled rating-bar graphics, with no digit or percentage attached to either. There is no comparison to Mistral's earlier OCR version, and no named individual, engineer, or spokesperson is credited anywhere on the page; the only identification given is the Mistral AI brand itself. Supported input file formats, page-count limits, and language coverage are likewise not stated.

Key facts

  • Mistral AI put OCR 4.1 into Public Preview on July 16, 2026, tagging it a "Premier" model reachable under three identifiers: mistral-ocr-4-1, mistral-ocr-4, and mistral-ocr-latest.
  • New in this version: native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores.
  • Standard output costs €3.5 per 1,000 pages; annotated output costs €4.38 per 1,000 pages, about 25% more.
  • BBox Extraction, OCR, and structured annotations are all served from the /v1/ocr endpoint; batch processing runs separately through /v1/batch.
  • The page publishes no accuracy, speed, or throughput figures and makes no comparison to the prior OCR version; only the Mistral AI brand is credited, with no named author or spokesperson.

Why it matters

OCR 4.1 pushes Mistral's document-parsing service beyond flat text extraction toward structured, geometry-aware output. Paragraph-level bounding boxes locate exactly where each piece of text sits on a page, structural block labels classify what kind of element it is, and block-level confidence scores indicate how much to trust each individual extraction rather than the page as a whole. For a pipeline built on top of OCR output, a retrieval system, a form-processing workflow, or a document archive, that combination is what lets low-confidence blocks get routed to a human reviewer instead of accepting an entire page uncritically.

Who it affects

Developers and companies already building on Mistral's Document AI stack, or evaluating OCR services for products that ingest scanned or PDF documents at scale (forms, contracts, invoices, archival records) are the direct audience. Anyone calling the API through /v1/ocr for BBox Extraction, plain OCR, or structured annotations, or through /v1/batch for bulk jobs, is working with the new version now that it is in Public Preview.

How to use it

OCR 4.1 is reachable now, in Public Preview, under three model names: mistral-ocr-4-1, mistral-ocr-4, or mistral-ocr-latest. BBox Extraction, OCR, and structured annotations all go through the /v1/ocr endpoint; batch jobs use the separate /v1/batch endpoint. Standard output costs €3.5 per 1,000 pages; requesting annotated output costs €4.38 per 1,000 pages, about 25% more.

How solid is it

Every fact here comes from Mistral's own product documentation page, the primary source for its own service, so specifics like the pricing, the endpoints, and the model identifiers are as reliable as vendor documentation gets. What is not demonstrated is performance. The page's "Speed" and "Performance" rows render only as filled rating-bar graphics, with no digit or percentage attached to either, and no independent benchmark is cited. There is no comparison to the prior OCR version, so there is no way to gauge from this source how much has actually changed, and no named individual, engineer, or spokesperson is credited anywhere on the page.

Risks and caveats

Public Preview status means the service can still change before it reaches general availability, and Mistral does not state supported input file formats, page-count limits, or language coverage anywhere on this page. Annotated output, one of this release's headline features, costs about 25% more per 1,000 pages than plain OCR. With no accuracy or throughput numbers published, anyone evaluating OCR 4.1 for a document-processing pipeline has no vendor-supplied benchmark to check it against.