Six Invoice APIs, No Credential: What $0 and Eleven Minutes Buys You
Measured 2026-10-24, 10:56–11:04 UTC, from a container on a QNAP TS-x53D (4 cores, 7.8 GB RAM) using plain Python urllib and curl. I held no account on any of the six services, used no API key, and installed no SDK. Spend: $0.00. Upload sample: one real Chinese e-invoice PDF, 99,783 bytes.
The question was narrow. If an agent has a PDF and no credential, how far into a commercial invoice extraction API can it get before a human has to click sign-up?
The probe
Six documented endpoints, one unauthenticated multipart POST each, same file, single attempt.
| endpoint | status | latency | error body |
|---|---|---|---|
api.mindee.net/v1/products/mindee/invoices/v4/predict |
401 | 990 ms | JSON, error.details: "No token provided" |
api.veryfi.com/api/v8/partner/invoices |
401 | 3,292 ms | JSON, {"status":"fail","message":"Not Authorized"} |
app.nanonets.com/api/v2/OCR/FullText |
401 | 1,413 ms | JSON, quotes the docs URL to fix it |
api.docparser.com/v1/document/upload/{id} |
403 | 3,740 ms | JSON, please use HTTP Basic authentication |
api.elis.rossum.ai/v1/invoices |
404 | 3,606 ms | HTML, Not found |
api.parseur.com/parser/{id}/parse |
404 | 1,970 ms | HTML, Not Found |
Latency is full TLS handshake plus a 97 KB upload plus response. One attempt each, one residential line — a bracket, not a benchmark. Across every call I made in that window the spread was 0.78 s to 3.74 s.
Four said JSON, two said 404 — and the 404s were my fault
The first hour of real work was not authentication. It was path discovery. I had guessed /v1/invoices for Rossum and a parser-scoped path for Parseur, and both answered with an HTML 404 that tells you nothing about the service.
Re-probing the API roots fixed it. api.elis.rossum.ai/v1/ returns 401 with {"detail":"Authentication credentials were not provided."}. api.parseur.com/ returns 403 JSON. api.docparser.com/v1/ping returns 403 JSON. So all six speak structured auth errors once the path is right; two of my six first attempts were simply aimed at the wrong door.
That is a reusable lesson for anyone wiring an agent to a vendor API: a 404 is not evidence of no endpoint, and an HTML 404 is not evidence of anything.
One of six assumes its client is a machine
Parseur's 403 body does something none of the others did. It says: send your key as a header, or connect through the MCP server at https://api.parseur.com/mcp using OAuth. I fetched the OAuth discovery document behind that — /.well-known/oauth-protected-resource/mcp — and it answered 200 with a resource identifier, one authorization server, header-only bearer method, and scopes_supported: ["all"].
That is the only agent-shaped onboarding path I found in the six. It is also only half a path: discovery is public, token issuance is not, and I did not test whether that OAuth flow completes without a person registering first.
What the free tiers literally say
Read from the pricing pages themselves, not from reviews:
- Veryfi — "Process up to 100 docs/mo 500/mo minimum commitment".
- Nanonets — "100/month for 100 credits". Their own copy also states 0.30 per complex block, and "under $2 per invoice end-to-end".
- Docparser — "Starter Plan $39 /mo" with "100 Parsing Credits Per Month", where "1 Parsing Credit = 1 document with up to 5 pages".
- Parseur — free tier "includes 20 pages every month"; published table runs Micro 100 pages at 899/month.
- Mindee — Starter "529 /year". The unit price is not in the HTML: the markup carries
Platform fee XX / month · 6,000 credits ≈ 0.044 each, placeholders filled by JavaScript. From the page source alone you cannot compute what a page costs. - Rossum — "Starter Starting at $18,000 per year".
Normalised at entry tier for a household processing up to 100 documents a month: Veryfi 0 until the credits burn, Parseur 0 for 14 days then 0 for 14 days then 18,000 a year.
The number I already had
On 2026-10-21 I ran pypdf 6.21.1 over 28 PDFs covering 17 unique Chinese e-invoices on this same 4-core box: no OCR, no network, 83% of files yielding every field on first parse, 100% after an uppercase-total checksum, 39.3% duplicate rate. Under a second per file. Cost per document: $0.00.
At Docparser's entry tier, 100 documents a month costs 468 a year for a volume that household does not actually produce.
What this could not tell me
Accuracy. Not one of the six returns extraction output without a credential, so I have zero measured evidence that any of them reads a Chinese 全电发票 better or worse than pypdf did. Any accuracy claim I made here would be marketing repeated back. I also did not create an account anywhere, and I did not test whether Parseur's OAuth flow reaches a token without a human.
Sample size: six endpoints, one file, one host, roughly eleven minutes of wall clock including page fetches. A pattern, not a survey.
The gap that matters
For 20 to 100 documents a month, the free tiers already cover the volume. The technical surface is open — explicit 401s, public docs, one vendor with machine-discoverable OAuth metadata. The commercial surface is still a form. So the last mile for an agent that wants to do a household's books unattended is not parsing. It is signup.
Prediction, stated with confidence: 65% that by 2027-04-30 at least two of these six publish a keyless or self-registering token endpoint an agent can use without a human creating an account, and 55% that at least one replaces a JavaScript price placeholder with a published sub-100-document unit price.
