Mistral OCR 4
State-of-the-art document OCR that turns PDFs and scans into structured Markdown.
Context window
Not announced
Input / 1M tokens
4000
Output / 1M tokens
4000
Provider
Mistral AI
Billed per page, not per token: $4 per 1,000 pages via the standard API, $2 per 1,000 pages with the Batch API, and $5 per 1,000 pages through the Document AI layer. · Data verified 2026-07-05
Mistral OCR 4 is Mistral AI's document-understanding and OCR model, released June 23, 2026. It extracts text with bounding boxes, classifies typed blocks (titles, tables, equations, signatures), attaches per-page and per-word confidence scores, and outputs clean Markdown. It accepts PDF, DOC, PPT, and OpenDocument files, supports 170 languages across 10 language groups, and offers an optional Document AI layer for reshaping output to a custom JSON schema. It can run fully self-hosted for data-sovereignty needs.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Call the Mistral OCR endpoint with model id 'mistral-ocr-latest' via Mistral Studio / La Plateforme, or use it on Amazon SageMaker and Microsoft Foundry. Self-hosted deployment is available to enterprise customers on request.
Strengths
- ✓Top-tier OCR accuracy (OlmOCRBench 85.20, OmniDocBench 93.07)
- ✓Structured Markdown output with typed block classification and bounding boxes
- ✓Per-page and per-word confidence scores for reliable downstream use
- ✓170-language coverage across 10 language groups, including low-resource languages
- ✓Extremely low per-page cost and optional fully self-hosted deployment
Best for developers who...
When to choose it (and when not to)
Reach for Mistral OCR 4 when...
- →You need to convert PDFs, scans, or office docs into structured text/Markdown
- →You require high-accuracy multilingual document parsing at low cost
- →You need tables, equations, and layout preserved with confidence scores
- →You must keep documents in-house (self-hosted deployment)
Look elsewhere if...
- ✕You need a general-purpose chat or reasoning model (this is OCR/document-only)
- ✕Your input is plain text with no document/image structure to extract
- ✕You need conversational tool-use or code generation
- ✕You need real-time streaming chat responses
How to use it
- ›Send high-resolution scans/PDFs for best extraction accuracy
- ›Use the Document AI layer with a JSON schema when you need typed, structured fields
- ›Use the Batch API to halve cost ($2 vs $4 per 1,000 pages) for large jobs
- ›Inspect per-word/per-page confidence scores to flag low-confidence regions for review
Quickstart
Pythonfrom mistralai import Mistral
client = Mistral(api_key="YOUR_MISTRAL_API_KEY")
resp = client.ocr.process(
model="mistral-ocr-latest",
document={
"type": "document_url",
"document_url": "https://example.com/invoice.pdf",
},
)
print(resp.pages[0].markdown)Uses the official 'mistralai' SDK OCR endpoint. Billing is per page processed, not per token.
API model id: mistral-ocr-4-0
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| OlmOCRBench | 85.20 | Top score reported in Mistral's OCR 4 announcement. |
| OmniDocBench | 93.07 | Reported in Mistral's OCR 4 announcement. |
| Human preference (head-to-head) | 72% win rate | 72% win rate against competing OCR systems in human preference evaluations, per Mistral. |
| Human evaluation (blind) | 72% | Independent annotators preferred OCR 4 in 72% of comparisons across 600+ documents |
| Human Evaluation (Blind Test) | 72% win rate | Independent annotators preferred OCR 4 over every competing system tested across 600+ documents in 12+ languages |