Mistral OCR.

Processing thousands of pages with equations or multilingual text? Need fast, cheap OCR via API without managing infrastructure? Mistral OCR outputs Markdown at $0.001/page.

We tested it.

Tested on December 17, 2025: 9-page PDF processed in 9.04 seconds, 34,656 chars output. Download output

The four metrics that matter.

Mistral's self-reported numbers.

Source: Mistral's internal benchmarks. Independent verification pending.

Mixed results in the wild.

• Koncile.ai: 98.75% transcription accuracy but 27.5% missing data in structured extraction
• Docsumo: "Even with moderately clean documents, it often missed key data blocks or misinterpreted layout structures"
• Parsio.io: Fast and cheap, but less robust than enterprise solutions for complex layouts

We recommend testing on your specific document types before production deployment.

Three steps to first parse.

pip install mistralai

export MISTRAL_API_KEY="your-api-key"

from mistralai import Mistral
client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
# Process a PDF from URL
ocr_response = client.ocr.process(
    model="mistral-ocr-latest",
    document={
        "type": "document_url",
        "document_url": "https://arxiv.org/pdf/2201.04234"
    }
)
# Get markdown output
for page in ocr_response.pages:
    print(page.markdown)

Process local files.

import base64

def encode_file(file_path):
    with open(file_path, "rb") as f:
        return base64.b64encode(f.read()).decode('utf-8')
client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
# Process local PDF
base64_pdf = encode_file("invoice.pdf")
ocr_response = client.ocr.process(
    model="mistral-ocr-latest",
    document={
        "type": "document_url",
        "document_url": f"data:application/pdf;base64,{base64_pdf}"
    }
)

Pricing comparison.

Mistral OCR vs Docling.

Same document: Docling paper (arxiv:2408.09869), tested December 17, 2025

Mistral is ~4x faster but costs money. Docling is free but requires local compute. Download test data

Fit for purpose.

Continue reading.