What Mistral OCR 4.1 actually is

Mistral OCR 4.1 is a public preview OCR model. The model card is dated July 16, 2026. The id is mistral-ocr-4-1. Mistral’s changelog that day says mistral-ocr-latest and mistral-ocr-4 now point to it.

It is not the first version that returns layout. OCR 4 (mistral-ocr-4-0) shipped June 23, 2026 as GA, with paragraph bounding boxes and structural block labels. OCR 4.1’s documented API addition is block-level confidence.

The job of the model is to turn a page into text plus where each region sits, what kind of region it is, and how sure the model is. Plain OCR that dumps a single string throws that away.

What OCR 4 already returned

Set include_blocks=True. Each page gets a blocks array in reading order. A block has type, top_left_x, top_left_y, bottom_right_x, bottom_right_y, and content.

The processor docs list these types:

text
title
list
table
image
equation
caption
code
references
aside_text
header
footer
signature

That split is what you use downstream. Drop header and footer from a search index. Keep title with the following text. Send table to a different parser. Attach caption to the nearby image. signature holds a transcribed name when the model can read it, otherwise an empty string.

Bounding boxes are how you rebuild column order or highlight the source region. They do not, by themselves, prove the reading order is the one a human would pick. Test that on your own two-column PDFs.

Block confidence

OCR 4 already scored pages and words. On July 16, the changelog added "block" to confidence_scores_granularity:

• page — page-level scores only
• block — page-level and block-level scores
• word — page-level and word-level scores

A block score can include average_content_confidence_score, minimum_content_confidence_score, and block_type_confidence_score. Content scores can be null on blocks with no text, such as images.

That is the hook for a review queue: high-confidence blocks go through; low-confidence ones wait for a person. A low score is a routing signal. It is not a certificate that the text is wrong, or that a high score is right.

How to call it

The processor lives at /v1/ocr. mistral-ocr-latest currently resolves to 4.1. Pin mistral-ocr-4-1 if you do not want the alias to move under you.

ocr_response = client.ocr.process(
    model="mistral-ocr-4-1",
    document={
        "type": "document_url",
        "document_url": "https://arxiv.org/pdf/2201.04234",
    },
    include_blocks=True,
    confidence_scores_granularity="block",
)

The URL has to be public, or you send a Base64 file or an upload. Images use image_url. The processor lists PDF, PPTX, and DOCX on document_url, and PNG, JPEG/JPG, and AVIF on image_url. You can limit pages. Tables can come back as markdown or HTML via table_format. Header and footer extraction is a separate flag pair, available on OCR 2512 and newer.

What to use it for

Mistral’s OCR 4 post frames the structured output as an ingestion layer for enterprise search, RAG, and agents: form filling, invoices, compliance checks. The same blocks are what you would store, chunk, or hand to a later model.

That is a pipeline design, not a benchmark. Run 20 to 50 pages of your reports, contracts, and scans. Check character accuracy, column order, table grid, caption bleed, block labels, and whether real errors actually land on low confidence. Official scores are directional. Mistral says the same about OlmOCRBench and OmniDocBench on the OCR 4 post.

Price and batch

The OCR 4.1 card lists $4 per 1,000 pages and $5 per 1,000 annotated pages. Recheck pricing before you budget.

/v1/ocr is a supported Batch endpoint. Batch processing is billed at a 50% discount to the standard rate. The June 23 OCR 4 post put that at $2 per 1,000 pages. Use it for archives and index jobs, not a UI that needs the page back immediately.

OCR 4.1 is still a preview. OCR 4.0 remains the GA card. If you need the latest aliases, you are on 4.1 today. If you need the GA line, pin mistral-ocr-4-0 and read what that card still includes.

Source cutoff: OCR 4.1 model card, changelog (June 23 and July 16, 2026), OCR processor docs, pricing, batch processing, and the OCR 4 launch post, as retrieved August 23, 2026.