A page is a page, and that is how we bill it
PDFs, scans, photographs of paperwork, handwriting. Munito AI turns them into Markdown and into the fields your application expects, on European infrastructure, with zero data retention by default.
The bill is a page count. Most platforms run OCR through a vision model and bill you per token, which makes a document budget a guess: a dense page and an empty page cost the same to read and very different amounts to pay for.
What it costs
| Route | What it does | Per page |
|---|---|---|
ocr-fast | Reads the text layer of a PDF, with no model | $0.0022 |
ocr-auto | Picks the model for each page | $0.0044 |
| PaddleOCR-VL | A vision model for the hard pages | $0.0066 |
A thousand pages through ocr-auto costs $4.40. A page with no text layer falls to a model; a page with one does not, which is why the fast route is half the price.
From a scan to a row in your database
| Workload | How |
|---|---|
| PDF to Markdown | Send the file, read the Markdown, keep the page order |
| Invoice extraction | Ask for a schema, get the fields, with the page each one came from |
| Tables | Structured output turns a table into rows, not into prose |
| Mixed bundles | One request reads a bundle of scans, photographs and born-digital pages |
Pair it with structured outputs and the response matches your schema every time, so the next step is an insert and not a parser.
The PDF to Markdown recipe and the invoice extraction recipe are working starting points.
Documents are the sensitive kind of data
An invoice carries a counterparty, a contract carries terms, a medical letter carries a patient. That is why the deployment question arrives earlier here than anywhere else.
Run on European infrastructure that Munito AI operates, inside your own cloud account, or on-premises. Munito AI keeps no page and no output by default, and trains on nothing you send. The DPA applies with no signature, at legal.munito.ai/dpa.
For the regulated versions of this workload, see Financial services, Legal and Healthcare.
Size the job before you commit
Run one representative batch, not a clean sample. Count the pages your pipeline actually fails on: rotated scans, two-column layouts, stamps over text, handwriting in the margin. Measure field accuracy against records your team already keyed, and the minutes each document saves.
A backlog is a batch job, not a live one. The batch tiers cut the price of work that can wait. See Service tiers.
What we do not do
We do not store your documents for you. A retrieved OCR result lives for 30 days only if your request asked to keep it; otherwise nothing is kept. We run no human review and no content moderation over your pages.
Start
- Read the Document AI guide
- See the per-page prices
- Extract invoice fields
- Talk to an engineer about your documents and your accuracy bar