AWS Textract Pricing: Amazon Textract Cost

Last updated August 2026.

AWS Textract costs 0.0015 dollars per page for plain text detection, 0.01 dollars per page for receipt and invoice extraction through AnalyzeExpense, and 0.05 dollars per page for form key value pairs. Rates drop at volumes above one million pages a month. There is no monthly platform fee and no minimum spend, so the bill is purely per page.

Every figure below was read from the AWS Price List API for the US West (Oregon) region in August 2026, not copied from a summary. AWS does not publish these rates in a greppable table on its pricing page, which is why so many third party posts quote stale or partial numbers.

AWS Textract pricing per page, US West (Oregon)

Textract bills by API and by feature type. The feature you request is what sets the rate, so the same page can cost anywhere from a fifth of a cent to seven cents depending on what you ask for.

For receipts and invoices, AnalyzeExpense at one cent a page is the meter that matters. It is purpose built for expense documents and returns named fields such as vendor, date, subtotal, tax, and total, along with line items. Using the cheaper DetectDocumentText meter for receipts is a false economy: you save 0.0085 dollars a page and inherit the entire job of working out which number on the page is the total.

The combination pricing nobody mentions

Here is the detail that changes real bills. If you request several AnalyzeDocument feature types in a single call, AWS charges a bundled rate that is lower than the sum of the parts. The saving is a consistent one cent per page on every combination.

The practical rule: if you need both forms and tables off the same document, ask for them in one call rather than two. Two separate calls bill both meters in full and put the page through twice. At a million pages a year that single implementation detail is ten thousand dollars.

Async costs exactly the same as sync

Every Textract meter is priced identically whether you call the synchronous or the asynchronous version. DetectDocumentText is 0.0015 dollars a page either way, AnalyzeExpense is a cent either way. That matters because the synchronous APIs cap out at 10 MB and a single page for PDF and TIFF, while the async APIs accept 500 MB and up to 3,000 pages. There is no price penalty for using async, so for anything multi page, use it.

The tradeoff is operational rather than financial. Async means submitting a job, polling for completion, and handling the cases where a job fails or a notification never arrives. A batch that silently stops running is the classic failure mode here, and it tends to be discovered at month end when the numbers do not reconcile, so it is worth putting a simple heartbeat check on the job that is supposed to run. Sync calls fail loudly and immediately, which is genuinely easier to operate, and the reason plenty of teams keep single page documents on the sync path.

What 2,500 receipts a month actually costs on Textract

Take a bookkeeping practice running 2,500 single page receipts a month through AnalyzeExpense. That is 2,500 pages at one cent, so 25 dollars a month. Run the same volume through Azure Document Intelligence on its prebuilt receipt model at 10 dollars per 1,000 pages and you also land on 25 dollars. The two big clouds price receipt extraction almost identically, which means the API rate is not the thing that should decide between them.

What the 25 dollars does not cover is an AWS account and IAM setup, the SDK integration, async job polling, retry logic, mapping the JSON response into your accounting format, and someone to maintain it when a dependency updates. For an engineering team processing millions of pages, 25 dollars against a few weeks of build is obviously correct. For a two person practice that just wants receipts in a spreadsheet, the build is the whole cost and the API price is a rounding error. Our comparison of an OCR API against finished tools lays out where each option stops being the cheaper one.

Limits that quietly change your bill

A few Textract constraints have cost consequences worth knowing before you budget. Synchronous operations accept 10 MB and one page for PDF and TIFF, so a 12 page PDF has to go async. Async operations accept up to 500 MB and 3,000 pages. Textract supports English, French, German, Italian, Portuguese, and Spanish, and handwriting recognition is English only, so a document set with other languages needs a different plan. Minimum text height is about 15 pixels, which means low resolution phone photos can fail and get resubmitted, and a resubmitted page bills again.

Is AWS Textract cheaper than Azure or Google?

For receipt and invoice extraction the three big clouds are close enough that price should not decide it. Textract AnalyzeExpense is 0.01 dollars per page. Azure prebuilt models are 10 dollars per 1,000 pages, so also 0.01 per page, with the detail broken down on our Azure Document Intelligence pricing page. Google Document AI charges 0.10 dollars per count for its invoice and expense parsers, where one count covers up to ten pages, which works out to 0.01 per page on a full ten page document and considerably more on a one page receipt.

That last point is the real difference. Google bills in blocks of ten pages, so a single page receipt costs 0.10 dollars there against 0.01 on AWS and Azure. If your documents are mostly one and two page receipts, Google is roughly ten times the price. If they are ten page invoice packets, the three are level. Page count distribution, not the headline rate, is what decides the winner.

Does AWS Textract charge per page or per document?

Per page. A five page invoice submitted to AnalyzeExpense bills as five pages. This is the most common reason Textract estimates come in low: teams count documents, multiply by the rate, and then find the real bill is two or three times higher because their average document runs to several pages. Before you budget, take a real sample and measure the average page count rather than assuming one.

Is there a free tier for AWS Textract?

Yes, AWS includes Textract in its free tier for the first three months of a new account, with separate monthly allowances per API. The allowances differ by API and AWS revises them, so check the current AWS free tier page rather than trusting a number in a blog post. What matters for planning is that the free tier is time limited: it is for evaluation, not for running a workload, and after three months every page bills at the standard rate.

Does Textract pricing include any accuracy guarantee?

No. Textract returns a Confidence value between 0 and 100 on each detected block and publishes no accuracy rate for receipt or invoice extraction. No major OCR vendor does, which is why every 99 percent claim in this market traces back to a marketing page rather than a benchmark. We cover what the metrics actually mean, and how to measure field accuracy on your own documents, on our OCR accuracy page. Budget for a human review step on low confidence fields, because it is a real ongoing cost that never appears in a per page quote.

Should you build on Textract or buy a finished tool?

Build on Textract if you have engineers, high or growing volume, and documents beyond receipts and invoices. The per page economics are excellent at scale, the combination pricing rewards a well designed integration, and you keep everything inside your own AWS account.

Buy a finished tool if what you actually need is receipts and invoices as a spreadsheet, this month, without a sprint. That is the job receipt OCR software does: upload a batch in a browser, get vendor, date, tax, line items, and total in Excel or CSV, and skip the account setup, the SDK, and the JSON parsing entirely. If you are still mapping the landscape, our rundown of the best OCR software separates the four different product categories that all get sold under the same word.