Mistral OCR

Mistral OCR is the pick when you want a hosted, per-page OCR API that keeps document layout, without running a parsing model yourself. Send it a PDF, an Office document or an image, and it returns per-page Markdown plus structured JSON: text, tables, images, hyperlinks, headers and footers, bounding boxes and confidence scores. The current model is OCR 4.1 (mistral-ocr-4-1). Document AI, Mistral's layer on top, adds structured extraction (annotations) billed per annotated page.

Highlights

• OCR 4.1 is current. Released July 16, 2026 and generally available since August 31, 2026; mistral-ocr-latest and mistral-ocr-4 both point to it. OCR 3 (mistral-ocr-2512) is still listed as active.
• Layout blocks in reading order. include_blocks returns paragraph-level bounding boxes, each with one of 13 structural labels such as title, table, equation, code, footer or signature.
• Confidence scores. Set confidence_scores_granularity to page, block or word, then route low-confidence regions to human review.
• Tables and page furniture. table_format returns tables as Markdown or HTML, extract_header and extract_footer pull running headers and footers out of the body text, and hyperlinks come back as their own output.
• Flexible input. Pass a public URL, base64 data or a file uploaded through the Files API. The pages parameter accepts ranges such as "0-5".
• Batch and cloud routes. /v1/batch handles bulk jobs, the model is also on Amazon SageMaker and Microsoft Foundry, and enterprise customers can self-host it in a single container.

In an AI-assisted workflow

Mistral's official SDKs are mistralai for Python and @mistralai/mistralai for TypeScript, both Apache-2.0. After pip install mistralai, this example from the Python SDK repo parses a PDF by URL and prints the JSON response:

import json
import os
 
from mistralai.client import Mistral
 
MISTRAL_7B_PDF_URL = "https://arxiv.org/pdf/2310.06825.pdf"
 
 
def main():
    api_key = os.environ["MISTRAL_API_KEY"]
    client = Mistral(api_key=api_key)
 
    # Using an URL
    pdf_response = client.ocr.process(
        document={
            "document_url": MISTRAL_7B_PDF_URL,
            "type": "document_url",
            "document_name": "mistral-7b-pdf",
        },
        model="mistral-ocr-latest",
        include_image_base64=True,
    )
 
    # Print the parsed PDF
    response_dict = json.loads(pdf_response.model_dump_json())
    json_string = json.dumps(response_dict, indent=4)
    print(json_string)
 
 
if __name__ == "__main__":
    main()

The example uses the mistral-ocr-latest alias; switch to mistral-ocr-4-1 once the output feeds a production index. From Claude Code, prompts like these turn it into an ingestion step:

Parse every PDF in ./contracts with mistral-ocr-4-1 and table_format set to html.
Write each page's markdown to its own file and keep the JSON response next to it.
 
Re-run page 3 with confidence_scores_granularity set to block and list the
lowest-confidence blocks so I can check them by hand.

OCR output is only half of ingestion. Tune chunk sizes on the Markdown with the chunking-strategy-optimizer skill before you embed it. When you need specific fields from invoices or forms rather than full pages, compare Document AI annotations with the schema-first vision-model approach in the multimodal-document-extractor skill.

How it compares

Mistral OCR sits with the managed APIs in our document parser roundup. Its closest siblings:

Choose Mistral OCR over LlamaParse when you'd rather budget against a flat per-page rate than map credit tiers to document types. Choose Reducto when you want parsing, extraction, splitting and classification as one platform, with VPC or air-gapped on-prem deployment. If documents can't leave your network and you have no enterprise contract, Docling and Marker run locally.

Good to know

Plans, as of September 2026 from Mistral's pricing and model pages: OCR 4.1 costs $4 per 1,000 pages for OCR and $5 per 1,000 annotated pages for Document AI. Mistral's June 2026 OCR 4 launch post advertised batch jobs at a 50% discount.

The service and the model are proprietary. The OCR weights are not open, and fully self-hosted deployment goes through an enterprise agreement. mistral-ocr-2505 (retired May 31, 2026) and mistral-ocr-2503 (retired December 31, 2025) are gone, so migrate any pipeline still pinned to them. Before committing, run a sample of your own documents through it, especially scans and non-English files, and check the Markdown and tables by eye. For how dedicated OCR models compare with general vision-language models on documents, see VLM OCR for documents.