How doc.cheap compares
Every comparison page is written by the winner, so here is the deal: the prices below are the vendors' own published list prices, the units are named because they differ, the things we have not measured are marked as not measured, and there is a section on what each of the others does better than we do. If a row looks unfavourable to us, it is because it is.
Published prices read on .
The table
| Price per document | Billed when the scan fails? | First call without an account? | Encoded zone + printed fields | Recognition latency | MCP server? | |
|---|---|---|---|---|---|---|
| doc.cheap | $0.01 flat, any volume | No — a failed recognition is free | Yes — public sandbox key, 10 per IP | Both, compared field by field | p50 ≈275 ms, measured in production | MCP server, three tools |
| AWS Textract (Analyze ID) source | $0.025/page to 100k/mo, then $0.01/page (us-west-2) | Not published as a no-charge case | No — AWS account and IAM first | Extracts ID fields; zone/print comparison not published | Not measured by us | No MCP server published |
| Google Document AI (passport parser) source | $0.10/document | Not published as a no-charge case | No — GCP project and billing first | Parser output; zone/print comparison not published | Not measured by us | No MCP server published |
| Mindee source | Credits from $0.05/document, +20% on overage | Not published as a no-charge case | No — account first | Extracts document fields; comparison not published | Not measured by us | No MCP server published |
| PassportEye (open source) | Free | — runs on your machine | Yes — it is a pip install | Encoded zone only | Depends on your hardware and images | No |
Prices are the vendors' published list prices read on 2026-09-21; Textract's figure is the us-west-2 rate and AWS prices differ by region. Re-check all of them before relying on this table — this is the kind of page that rots.
We do not quote a price for Azure's document service: its pricing page did not render for us, and we will not publish a number we have not read.
Read the units before you read the numbers
Some vendors bill by the page, we bill by the document. A passport data page is one page and one document, so for that case the two units coincide — but a multi-page document is one document here and several pages there, and the comparison stops being like-for-like immediately. Where a vendor's unit is the page, the table says so.
The second difference is what a failure costs. Here, a scan that finds no document, cannot read the image, or cannot identify the type is answered and not charged, and a billed flag on every response says which it was. We have not found a published equivalent rule at the others; treat their cells as “read the current terms”, not as “they charge you”.
Where each of them beats us
AWS Textract and Google Document AI — breadth
Both read far more than identity documents: invoices, receipts, forms, tables, handwriting, arbitrary layouts, at scale, in a dozen languages, wired into the rest of their cloud. We read identity and travel documents. If your pipeline needs one vendor for passports and supplier invoices and scanned contracts, they cover the whole job and we cover a third of it. Their regional footprint and their compliance paperwork are also far ahead of a company our size.
Mindee — custom models
If your document type is not in anybody's catalogue — your own internal form, a regional permit, a carrier's waybill — Mindee lets you train a model for it. We do not. That is a real capability gap, not a positioning difference.
PassportEye — free and offline
It costs nothing, it runs on your own machine, and no image ever leaves the building. For clean scans of passports where the encoded zone is all you need, it is the right answer and we are the more expensive one. Its limits are equally real — the printed side is not read, phone photographs need your own pre-processing, and you own a tesseract install in production — but “free and local” is a genuine advantage and we will not pretend otherwise.
Where we are ahead
- Price, for the identity-document case. 1¢ per document. Same price for everyone — from 1 to 100 million documents. Beyond 100,000 pages a month Textract's rate meets ours rather than exceeding it.
- A failed scan is free, and the response tells you.
- You can call it before you have an account. No cloud project, no IAM role, no billing setup, no card: a sandbox key printed in the documentation.
- Both zones, compared. Every field is read from the machine-readable zone and from the printed visual zone separately, each with its own confidence, so a disagreement between the encoded date of birth and the printed one is visible. Check digits are verified against ICAO Doc 9303.
- Native script beside the Latin. Values in Greek, Arabic, Han, Hangul or Thai come back in their own script, tagged with their language, not flattened to ASCII.
- Nothing stored. The image is never written to disk, results are kept only for the retain_hours you ask for, and that defaults to zero.
- A measured latency in the contract. A p50 of about 275 ms of recognition time in production, every response carrying its own timing split, and the live figure published on the status page.
- An MCP server, so an assistant calls the API directly with three tools and no SDK in between.
Same price for everyone — from 1 to 100 million documents. The first call needs no account, the billing rules are published in full, and the MCP server is the same recognition reached by an assistant.
Which one to pick
- You need passports and ID documents, cheaply, with both zones read
- this one. Try it before you decide; the first call needs no account.
- You need one vendor for every document your business touches
- Textract or Document AI. Breadth is what they sell, and they sell it well.
- Your document type does not exist in any catalogue
- Mindee, and train it.
- The image must never leave your network, and the encoded zone is enough
- PassportEye, plus a real image pre-processing pipeline.