AWS Textract cost: per page, by the analysis you run

Quick answer

Textract bills per page processed, at rates that rise with the analysis: plain text detection is cheapest, while forms, tables, and query features cost more per page, and combining them adds up. So the levers are running only the analysis each document type needs, and not extracting forms and tables from documents that only need text.

Amazon Textract extracts text, forms, and tables from documents. It is priced per page, with the rate depending on which analysis features you request. The common overspend is running the advanced, more expensive analysis on documents that only need simple text extraction.

Per page, by feature

Feature	Relative cost
Text detection (OCR)	Lowest per page
Forms analysis	Higher per page
Tables analysis	Higher per page
Queries / signatures / layout	Additional per page

Plain text detection is the cheapest per page. Extracting form key-value pairs, tables, or using query features each costs more, and requesting several together bills for the combination. So a page processed with forms and tables analysis costs significantly more than the same page for text only.

Match analysis to document type

The saving is running only the analysis each document actually needs. A document you only need the text from should use text detection, not forms-and-tables analysis. A form should use forms analysis. Routing documents to the right analysis by type, rather than running full analysis on everything, avoids paying advanced-feature rates for simple extraction.

Controlling Textract cost

Use the cheapest analysis that meets each document type's need, avoid running forms and tables analysis on text-only documents, cache or store results so you do not reprocess the same document, and pre-filter blank or irrelevant pages. As with any per-unit service, matching the request to the need is the optimization, the same discipline as Rekognition.

FAQ

How is AWS Textract priced?

Per page processed, at rates that rise with the analysis: plain text detection (OCR) is cheapest, while forms analysis, tables analysis, and query, signature, and layout features each cost more per page. Combining features adds up, so a page with forms and tables analysis costs significantly more than text-only.

How do I reduce Textract cost?

Run only the analysis each document type needs, use cheap text detection for text-only documents rather than forms-and-tables analysis, route documents to the right analysis by type, cache results to avoid reprocessing, and pre-filter blank or irrelevant pages. Matching the request to the need is the main lever.

Why does forms and tables analysis cost more in Textract?

Because extracting structured form key-value pairs and table cells is more complex than plain text detection, so those features are priced higher per page. Requesting several features together bills for the combination, so full analysis on every document is far more expensive than targeted text extraction.

Should I run full Textract analysis on every document?

No. Running forms and tables analysis on documents that only need text extraction over-pays for features you do not use. Route documents to the appropriate analysis by type, text detection for text-only, forms analysis for forms, so each page is processed at the cheapest sufficient rate.

Does Textract charge per document or per page?

Per page. A multi-page document bills for each page processed, at the rate for the analysis requested. So document length and the analysis features together determine cost, and pre-filtering unnecessary pages (blank pages, irrelevant sections) reduces the page count billed.

Does C3X estimate Textract cost?

Textract cost is driven by pages processed and analysis features, which are usage inputs. C3X prices the surrounding infrastructure, and you model your document page volume and the analysis each type needs to estimate the per-page charges.

What to do next

Estimate the infrastructure around your document pipeline before you build it. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.