Mistral OCR 4.1

Who Should Read

A backend developer weighing an OCR API over Tesseract for PDF and scanned-document RAG or extraction pipelines

Core Mechanics

• OCR 4.1 is the latest Document AI OCR
  • The model ID is mistral-ocr-4-1.
  • It is listed as GA on July 16, 2026.
• Paragraph-level coordinates come built in
  • It natively extracts paragraph-level bounding boxes.
  • Structural block labels come with them.
• Every block gets a confidence score
  • It provides block-level confidence scores.
  • The page gives no score range or usage guidance.
• Pricing is per page
  • Standard pages are listed at $4 or €5 per 1000 pages.
  • Annotated pages are $5 per 1000 pages.
• Endpoints are /v1/ocr and /v1/batch
  • BBox extraction, OCR and structured annotations use /v1/ocr.
  • Batching uses /v1/batch.

Community Reactions

• Weak on edge cases, per one user
  • A user OCRing a book scan with ligatures, Fraktur and sub/superscripts saw nothing special.
  • They say OpenAI pro models do better, though even top models struggle with intricate text.
• Cheap and fast on simple docs, some say
  • One user converting PDFs to EPUB says it tagged headers, titles and footnotes on a Bleak House chapter every time.
  • Right-margin line numbers were tagged as aside text 3 of 5 times.
  • They add that no prompting is needed and batch mode is 50% off.
• Many found the price too high
  • Commenters say about 3.5 per 1000 pages must be far better than Tesseract to be worth it.
  • One runs their own pipeline on rented GPUs at 0.05-0.1 USD per 1000 pages, about 0.8s per page.
  • Another says Mistral raises the price each release.
• Alternatives and trust concerns
  • One user runs NuExtract locally on a single RTX 4090 for bank statements and finds it beats Tesseract and generic vision models.

How to Apply

• Compare on 20 of your own documents
  • Run the same PDFs through your current Tesseract (OcrMyPDF) setup and /v1/ocr.
  • Mix in tables, footnotes and multilingual pages.
• Use bboxes and block labels for chunking
  • When indexing for RAG, split chunks by header, body and footnote using block labels and bboxes.
  • Highlight the source region with the bbox when citing answers.
• Route low-confidence blocks to review
  • Set a threshold on block confidence and send low blocks to human review or a second model.
  • Pick the threshold from your sample test results.
• Run non-realtime jobs through /v1/batch
  • Send bulk backfills to the batch endpoint.
  • Check the commenters' 50% discount claim against the official price list.

Try It

# Paste into Claude or ChatGPT: a prompt to plan an OCR evaluation before adopting

I want to use PDFs for RAG indexing. I currently use {{current OCR tool}}.
The candidate is Mistral OCR 4.1, which offers paragraph bboxes, structural block labels, block confidence scores, and $4 per 1000 pages.
My document types are {{document types: contracts/scanned books/table-heavy reports, etc.}}.

Create the following.
1. Criteria for picking 20 samples (include tables, footnotes, multilingual, low quality)
2. Accuracy metrics and scoring method
3. A procedure to set the block confidence threshold
4. A cost comparison table at {{monthly page count}} pages per month

Next step: paste the results from your samples and ask it to classify errors by type.

Terminology

Related Papers

• You could have invented transformers from memory

  Viewing linear, MLP and attention layers as memory stores: capacity goes ~64 to ~2,700 to over a million

• Pointing AI at archives found a forgotten meteorite, lost rhinos, and more

  Embedding search plus a cheap filter model surfaced unrecorded events in 400 years of archives

• EmbeddingGemma 2: An open, lightweight multimodal embedding model

  One open 740M model embeds text, images, audio and video into a single shared vector space

• RAG-NAROK: Retrieval-Aware Knowledge Corpus Poisoning in RAG with Source-specific Refutation

  Refutation documents targeting retrieved sources took top rank and evaded anomaly detectors up to 88.37%.

• Text, Pixels, or Both? Evaluating Input Representations for Multimodal Document QA

  Page images beat OCR text on accuracy but cost more; a question-only router cut latency 30%.

• Microsoft exec called AI scraping 'the largest theft of labor in human history'

  NYT lawsuit unredacted filings reveal Microsoft and OpenAI internal views on scraping

Related Resources

• https://docs.mistral.ai/models/ocr-4-1