Mistral OCR 4 with n8n: automating invoice and document processing

What Mistral OCR 4 Provides

OCR 4 has the versioned identifier mistral-ocr-4-0. The documented alias for tracking the current version is mistral-ocr-latest.

For each page, a response can contain:

• extracted Markdown;
• page dimensions;
• extracted images, where applicable;
• layout blocks when include_blocks is true;
• confidence scores when confidence_scores_granularity is page or word.

Blocks describe their type and coordinates with fields such as top_left_x, top_left_y, bottom_right_x and bottom_right_y. A workflow should not depend on a generic bbox field or a paragraph block type that is absent from the documented contract.

Mistral reports a score of 85.20 on OlmOCRBench and support for 170 languages. Vendor results are a useful reference, but they do not predict performance on your invoices, scans or tables. Test a representative document set before production.

Pricing and Hosting

OCR 4's published price is $4 per 1,000 pages, or $5 per 1,000 annotated pages. The price of a document therefore depends on its page count, before storage, orchestration and review costs.

Mistral also offers a self-managed, single-container deployment in its enterprise offering. This can address a hosting constraint, but it is not a free switch. Evaluate licensing, infrastructure, updates, monitoring and security requirements.

Price per page is not enough to compare products. Run the same document set through the shortlisted services and measure:

• text and table quality;
• correctly extracted business fields;
• the share of cases routed to review;
• latency on your actual volumes;
• total cost, including operations and errors.

A provider's nationality does not establish compliance by itself. Review the processed data, contract, processing region, retention, subprocessors and transfers for the actual use case.

Calling the API from n8n

The minimal workflow is:

[Receive document]
        -> [HTTP Request to Mistral OCR]
        -> [Field extraction and business checks]
        -> [Human review where needed]
        -> [Target system]

Configure an n8n HTTP Request node to send a POST request to https://api.mistral.ai/v1/ocr, using Bearer authentication and a JSON body.

For a PDF available at a URL:

{
  "model": "mistral-ocr-latest",
  "document": {
    "type": "document_url",
    "document_url": "https://storage.example/invoice.pdf"
  },
  "include_blocks": true,
  "confidence_scores_granularity": "word"
}

For a base64-encoded PDF, the documentation still uses document_url with a data URL:

{
  "model": "mistral-ocr-latest",
  "document": {
    "type": "document_url",
    "document_url": "data:application/pdf;base64,{{ $json.base64 }}"
  },
  "include_blocks": true,
  "confidence_scores_granularity": "word"
}

An image uses the image_url type instead. This distinction is easy to miss when the document starts as binary data in an n8n workflow.

If the PDF is behind a signed URL, give it enough validity for the request and avoid logging the URL. For confidential documents, also limit what n8n retains in execution history.

Extracting Fields Without Inventing Certainty

OCR produces text and document structure. It does not automatically produce an accounting-ready invoice. The next layer must turn the response into business data and reject inconsistent cases.

For a French invoice, checks can include:

• the format of the SIREN or SIRET and, where needed, verification against an appropriate source;
• presence of an invoice number;
• consistency between issue date and due date;
• reconciliation of line totals, VAT and the total due;
• currency and decimal separators;
• duplicate detection before creation.

A confidence score is not a universal probability that a field is correct. Calibrate it on your documents. Start with human review of all results, measure errors by field type, then define different thresholds if the observations support them.

For example, a readable company name may need only a light check, while a changed bank account or an amount that triggers payment should remain subject to stronger approval. Business risk matters more than one threshold applied to every field.

Three Reasonable Use Cases

Prefilling a Supplier Invoice

n8n retrieves an attachment from a dedicated inbox, calls OCR 4, then prepares a record with the supplier, invoice number, dates and amounts. An accountant compares it with the PDF before approval. The workflow publishes nothing if totals fail to reconcile or a duplicate is found.

Indexing Contracts

OCR can make scanned archives searchable. Access rights, retention and deletion still belong in the document system. The index must not become a copy available to everyone in the company.

Processing Expense Receipts

Receipt photos can be extracted and matched to a transaction. An unreadable date, missing currency or several candidate totals should route the receipt to review instead of forcing an answer.

What OCR Does Not Replace

OCR 4 does not replace an approved invoicing platform or a structured e-invoice format. From September 1, 2026, all VAT-liable businesses established in France must be able to receive electronic invoices. Large companies and mid-sized businesses must start issuing them on that date; SMEs and microbusinesses follow on September 1, 2027.

OCR remains useful for unstructured documents, archives and flows outside the scheme. For the timetable and preparation steps, see the article on electronic invoicing for French companies.

Sources Consulted

• Mistral AI, Mistral OCR 4 announcement, June 23, 2026.
• Mistral AI, OCR 4 model card, accessed August 13, 2026.
• Mistral AI, OCR API documentation, accessed August 13, 2026.
• Mistral AI, pricing, accessed August 13, 2026.
• French tax administration, electronic invoicing timetable, accessed August 13, 2026.

A Sensible First Version

The first useful version should not process every document. Start with one homogeneous family, keep the source PDF, extract a few fields, apply simple checks and provide a review queue.

Then measure errors and time actually saved. If the workflow is reliable, expand its scope. Kirako's AI agents and automation page describe how I can help design and connect it to your tools.