Skip to content

Document Processing Cost Estimator

Price a batch of documents through a text-extraction pipeline and through a vision model, and see which route is cheaper at your page count.

Cheaper route for 1,000 documents
$34.15

Text extraction, then the model. The other route costs $39.75 — a difference of $5.60 on this batch.

Pages in the batch
8,000
Text route — input tokens
4.80M
Text route — parsing / OCR
$16.00
Text route — total
$34.15
Vision route — input tokens
12.00M
Vision route — total
$39.75
Output tokens (both routes)
250,000
Output cost (both routes)
$3.75
Text route, per document
$0.03
Vision route, per document
$0.04
What this assumes: one model call per document, with all of its pages in a single prompt. If you call per page instead, the output tokens multiply by the page count and the gap between the routes narrows sharply. No retries, no re-processing of documents that failed validation, and no second pass for the ones the first pass got wrong — in a real extraction pipeline that second pass is a line item, and it is not on this page.

Two routes into the same model

A page can reach a model as text or as a picture, and the two are priced on entirely different meters. Extracted text costs whatever its token count is, typically a few hundred tokens for a page of prose. The same page as an image costs whatever the provider’s tiling formula says, which is a function of resolution and is frequently several times higher. Against that, the vision route deletes the parsing step and everything that can go wrong in it.

Which wins depends on your documents, and the crossover is real rather than theoretical. Clean digital PDFs with a reliable text layer favour extraction by a wide margin. Scans, screenshots, forms where position carries meaning, and anything with tables that a parser flattens into nonsense favour vision — not because it is cheaper but because the extraction route quietly produces wrong answers, and a wrong answer costs more than any token.

What the batch number hides

Per-document cost is the number to plan with, but the number that decides the architecture is cost per correct document. If one route needs a human to check ten percent of its output and the other needs a human on two percent, the difference in review time will dwarf the difference in tokens at any serious volume. Price both routes here, then measure the accuracy on fifty real documents before choosing.

Document Processing Cost Estimator · Multigrid