Mistral Document AI
Turns a PDF or an image into page-level Markdown in one call, with the exact fields you name pulled out by JSON schema in the same pass.
Where Mistral Document AI lands
No score or rank is published until the shared benchmark corpus covers Mistral Document AI — the note below carries the reason. Full method: how we measure.
The honest pitch
Reach for it when a document is the input and structure is the output: contracts, invoices, receipts, scanned reports — page-level Markdown with tables intact, plus a JSON object holding just the fields your schema names. Point it at an HTTPS link Mistral can reach, say which pages matter, and the cost is known before the call runs.
Look elsewhere when the file can't be reached over HTTPS — the document travels as a link, not an upload — or when you want a model to reason over the text and answer questions about it, which belongs on an inference route and prices per token rather than per page. Page selection is required and caps at 100 pages a request, so an unbounded archive in a single shot isn't the shape this route takes.
About Mistral Document AI
Mistral’s Document AI reads a PDF or an image and hands the document back as Markdown, page by page — headings, paragraphs and tables preserved in a shape an agent can work with instead of a wall of loose text. Add a JSON schema to the same request and the response also carries the fields you asked for: an invoice’s number and total, a contract’s parties and dates, a label for each figure it finds. Transcription and extraction land together, on one call.
Through NativePort it is a single fixed route — POST /mistral/v1/ocr, pinned to the mistral-ocr-4-0 model. The document or image travels as a URL: send an HTTPS link Mistral can reach, name the pages you want (at most 100 per request, zero-based), and the answer comes back exactly as Mistral returns it. The link does not have to be public — a signed, time-limited URL can work just as well, provided it is still valid when the call runs. Because the page selection is explicit, the price is known before the work starts: $4 per 1,000 pages for page Markdown, $5 per 1,000 pages once a request carries an annotation schema, at Mistral’s own published rates and never both on one call. Successful calls are metered per processed page, bounded by the page selection priced before work begins. The charge draws on the same NativePort balance as the rest of the catalog, and a job your balance can’t cover is refused instead of run.
Documents like these are usually the sensitive kind, so it matters where the text ends up on the way through. For this route, the gateway omits request and response bodies from its logs, so neither the link you send nor a line of extracted text is written there — its usage records hold the operation, the page count and the rate. What Mistral itself keeps on its own side is governed by Mistral’s terms, not by the gateway. Field-by-field request and response detail lives in the endpoint reference.
Reaching it through NativePort
One route. The gateway applies this route's documented request policy, injects the upstream credential and returns Mistral Document AI's response, metered against your balance at Mistral Document AI's real published usage price — request and response fields sit in the endpoint reference.
curl -X POST https://api.nativeport.ai/mistral/v1/ocr \ -H "Authorization: Bearer $NATIVEPORT_API_KEY" \ -H "Content-Type: application/json" \ --data '{"document": {"type": "document_url", "document_url": "https://example.com/contract.pdf"}, "pages": [0]}'