Pricing pages for document-extraction APIs are written to win a different comparison than the
one you actually need. Per-page rates hide minimums. Credit systems hide per-document math.
So here's the arithmetic for one concrete, boring workload: 1,000 single-page invoices a month, at each vendor's published July-2026 pricing. (I sell one of these APIs — disclosure
| API | Published rate | 1,000 invoices/mo | The catch |
|---|---|---|---|
| AWS Textract (AnalyzeExpense) | $0.01/page | $10 | |
| You build the pipeline: retries, failures, schema mapping. Every attempt bills, including failed ones. | |||
| Google Document AI (Invoice parser) | $0.01/page | $10 | |
| Same deal. (Their Bank Statement parser is $0.75/classified doc.) | |||
| Azure Document Intelligence (prebuilt) | $0.01/page | $10 | |
| Same again — the hyperscalers are interchangeable here. | |||
| LlamaParse | $1.25/1k pages | ~$1.25 | |
| It's a parser, not an extractor: structure back, not validated invoice fields. Budget your own extraction layer (and its LLM bill) on top. | |||
| Veryfi | $0.16/invoice | ||
| $160 — except no. Production starts at a $500/mo minimum buying <5k docs | |||
| At 1,000 invoices your effective rate is $0.50/doc, not $0.16. | |||
| Kynth Core (mine) | $0.08/invoice | $80 | |
| Charged only when extraction succeeds; no minimum, no commitment. |
1. What does a failure cost? The per-page APIs bill per attempt. If your docs are ugly —
photographed receipts, scanned faxes, tables that explode — your effective
per-successful-document rate is higher than the sticker. That's why "cost per correctly
extracted document" (price ÷ accuracy) is the only number worth comparing, and why I publish
a reproducible accuracy benchmark alongside this post — including the suites where the
hyperscalers beat me.
2. What does the minimum cost? Veryfi's per-doc rates look adjacent to mine until the
commitment line. Minimums are a bet the vendor makes you place on your own volume. Process
800 invoices a month and you're paying for 5,000.
3. Who owns the pipeline? $10/month at Textract is real if your engineering time is free.
The hyperscaler APIs hand you fields; everything around them — schema validation, retries,
normalization, the 3am alert when a new invoice layout breaks parsing — is yours.
Kynth Core is my product, so weight my row accordingly. What I can offer instead of
neutrality is reproducibility: the accuracy benchmark behind these claims is an MIT-licensed
harness (public datasets, pinned revisions, raw responses committed, fork-safe CI) — re-run
every number with your own keys, or re-score my committed responses with no keys at all.
Links: [benchmark repo](https://github.com/kyisaiah47/doc-extract-bench) ·
[live scores](https://api.kynth.studio/benchmarks) ·
[pricing sources](https://api.kynth.studio/guides/invoice-ocr-api-pricing).