Amazon Textract Pricing: AWS Cost Per Page 2026

The short answer: Amazon Textract charges per page, per feature, with no subscription and no minimum spend. In US West (Oregon) plain OCR is $1.50 per 1,000 pages, table extraction is $15.00 per 1,000, and form key value extraction is $50.00 per 1,000 for the first million pages each month. Those charges are additive, so a page sent through both tables and forms costs $0.065 rather than $0.050. The free tier lasts three months and does not include table extraction.

That paragraph answers the search, but it is not the whole bill. Most teams who budget from the rate card end up somewhere else, either because they enabled a feature they did not need or because the AWS line item was never the expensive part. Here is the full picture, with the numbers checked against the AWS pricing page and the Textract developer guide in August 2026.

Amazon Textract pricing table

Textract has no plans and no seats. You pick an API operation, and each page you send through it is billed at that operation's rate. Prices vary by AWS region; these are the published US West (Oregon) rates.

Two things stand out. Plain OCR is roughly ten times cheaper than table extraction, which is why a surprising number of pipelines run Detect Document Text and then reconstruct columns from word coordinates in their own code. And Layout costs nothing when you already asked for Tables, so there is no reason to leave reading order on the table if structure matters to you.

Does Amazon Textract charge separately for tables and forms?

Yes. Every feature you enable on a page is billed independently, and they add together. A single AnalyzeDocument call requesting both TABLES and FORMS costs $0.015 plus $0.050, so $0.065 for that page. Nothing warns you. The API accepts the list of feature types and returns one combined response, and the charges appear as separate line items later.

This is the single most common way a Textract estimate goes wrong, because the AWS code samples and most tutorials enable several features at once to show the full response shape. Copy one of those into production, run 20,000 pages a month through it, and you have a $1,300 bill where $300 would have done. Before you scale anything, look at what your parser actually reads out of the response and delete every feature type you never touch.

Is Amazon Textract free?

There is a free tier, but it runs for three months from account creation and it deliberately excludes the feature most people arrive for. The allowances per month are 1,000 pages of Detect Document Text, 1,000 pages of Analyze Document restricted to signature detection, 100 pages each of Analyze Expense and Analyze ID, and 2,000 pages of Analyze Lending. Table extraction is billed from the first page.

So the honest reading is that you can test OCR for free and you cannot test tables for free. It is a small amount of money, $1.50 for a hundred page trial, but it does mean the evaluation everyone actually wants to run starts on the meter.

How much does it cost to process 1,000 pages with Amazon Textract?

Depends entirely on which feature you call. Here is the same 1,000 pages through five realistic configurations.

Scale that up and the shape holds. Table extraction on 10,000 pages is $150 a month, on 100,000 pages it is $1,500, and past a million pages the rate drops to $0.010 and the economics get genuinely attractive. Textract is priced for volume, and at volume it is good value.

Are there hidden costs with Amazon Textract?

Nothing hidden in the dishonest sense, but the rate card is not the invoice. Four things sit alongside it.

S3 storage and requests. Any PDF longer than one page has to go through the asynchronous API, and asynchronous operations read the file from S3. So every multi page document gets uploaded, stored and eventually deleted, and all of that is billed separately. Individually trivial, and worth knowing about before someone asks why the storage line moved.

Compute around the call. Textract's asynchronous operations do not block, so something has to start the job, wait for the SNS notification or poll, then page through the results. In practice that is Lambda invocations, or a container that is running anyway, plus the CloudWatch logs it generates.

Retries on bad input. Textract rejects password protected PDFs and XFA based forms outright, and it will not read vertical text or characters under about 8 point at 150 DPI. Failed pages still cost you an operation if the failure happens mid document, and the retry costs another.

The engineering. This is the big one and it never appears on any pricing page. Textract returns JSON Block objects, not a spreadsheet. Somebody has to write and maintain the code that walks blocks, reassembles cells into rows, handles merged cells and tables that continue across a page break, and writes the result to a file. AWS publishes Python samples and the open source Textractor library will export a table with a to_excel call, which shortens the job considerably, but it is still code your team owns. If you are tracking what the whole thing costs rather than what AWS invoices, put that in the model, and keep an eye on where cloud and SaaS spend actually lands each month so a per page service does not quietly become a fixed cost nobody reviews.

Amazon Textract pricing vs Azure and Google

Buyers rarely price Textract alone. The two obvious comparisons are Azure AI Document Intelligence and Google Document AI, and all three land in the same range per page.

We are quoting the Azure figure from third party pricing summaries rather than Microsoft's own page, because that page renders its rate table dynamically and the numbers are not in the page source. Treat it as indicative and confirm in the Azure pricing calculator. The Azure free tier is the genuinely different item: 500 pages a month with no 90 day clock makes evaluation a lot easier than Textract's expiring allowance.

What none of the three change is the last column. All of them return structured JSON, and none of them hand you an .xlsx file. If the deliverable is a spreadsheet on somebody's desk, that gap is the project.

Per page billing versus per file

This is the comparison that decides most real purchases, and the rate card cannot show it. Textract bills by page. A converter aimed at business users generally bills by conversion or by subscription, which means document length stops mattering.

Take bank statements, which are the worst possible shape for per page pricing: repetitive, long, and nobody chose the page count. A 42 page statement is 42 billed pages of Tables, about $0.63. Fifty of those in a month is roughly 2,100 pages and about $31 of Textract charges, before any of the surrounding infrastructure. A flat plan at $9 a month covers the same work because it counts files, not pages.

Reverse the volumes and the answer reverses with them. At 200,000 pages a month Textract Tables is $3,000, which sounds like a lot until you notice nobody is clicking upload 200,000 times. At that scale you want the API, and the engineering cost amortizes into nothing. The break point sits lower than most people assume, somewhere around the volume where a person stops being able to handle the documents by hand.

What is the cheapest way to get tables out of a PDF with Textract?

Call Tables and nothing else, use the asynchronous API so you are not paying an operation per page on multi page files, and strip every feature type your parser does not read. If your documents are clean and ruled, test Detect Document Text at $1.50 per 1,000 first and see whether you can reconstruct columns from word geometry, because that is a tenfold saving when it works. It often does not work on borderless financial tables, which is exactly where Tables earns its price.

Also check whether you need Textract at all for the job in front of you. A one off analysis, a monthly reporting pack, or a folder of statements a bookkeeper has to reconcile are all tasks where the setup outweighs the saving. Our Amazon Textract alternative page walks through where the line falls, and a plain PDF table extractor handles the one off case in seconds.

Does Amazon Textract have a minimum spend or a commitment tier?

No minimum and no commitment. Textract is pure pay as you go, and the only volume discount is automatic: rates drop once you pass one million pages in a calendar month for most features, and 100,000 pages for Analyze ID. There is nothing to sign and nothing to forecast, which is a genuine advantage over vendors who quote annually and want a call before they show you a number.

That flexibility is the strongest argument for Textract in a spiky workload. If you process 400 pages in March and 90,000 in April, you pay for 400 and 90,000. A subscription charges you the same either way, and a quoted enterprise contract charges you for the peak you predicted.

How accurate is Amazon Textract on tables, and what does inaccuracy cost?

Textract handles clean, ruled tables well and reports merged cell relationships, which many extractors skip entirely. Borderless financial tables, multi line cells, and rows that continue across a page break are harder, and they are harder for every tool including ours. AWS does not publish a single accuracy figure for tables, and any vendor who does is quoting a number measured on their own document set.

The cost that matters is not the $0.015. It is the twenty minutes someone spends fixing a column that shifted, multiplied by every document. Run your worst file rather than your cleanest one during evaluation, then check the result properly: our guide to checking a PDF to Excel conversion for errors lists the five checks that catch almost everything, and cleaning up converted data in Excel covers what to do when something did shift.

Which Textract feature should I actually call?

Picking correctly here is worth more than any negotiation, because the gap between the right feature and the wrong one is often five times the price. A purchase order looks like a form and is usually answered by Tables. An invoice looks like a table and is usually answered by Analyze Expense at two thirds the price.

The bottom line on Amazon Textract pricing

Textract is cheap per page and expensive per project. If you are building document processing into software, the rate card is close to unbeatable and the absence of any commitment makes it easy to start. If your goal is a spreadsheet you can open, the $0.015 was never the number to evaluate, because the response is JSON and the work is in what happens after it.

Price both halves before you decide. Work out the AWS charge from the table above, then estimate the days of engineering time to turn blocks into files, and compare the total against a flat plan. For a lot of finance and operations teams the honest answer is that they need statements converted to Excel a few dozen times a month and never needed an API at all. For teams processing at real scale, Textract is the right tool and the per page rate is the least of the reasons.