Document AI pricing: the page rate is the smallest number on the bill

The short version

• Document AI services generally bill per page, and the rate depends on which extraction feature or processor you call.
• Plain text extraction is the cheapest tier; extracting forms, tables, invoices or custom fields costs more per page.
• Google's Document AI price list includes a $0.05 hourly hosting rate, which is $438 over a year of continuous deployment.
• The per-page API fee is rarely the largest cost. Human review, exception handling and integration usually cost more.
• Estimate cost by running a sample of your own documents and counting pages, not documents.

How is document AI priced?

Document AI is priced per page, and the rate depends on the feature you call. Reading plain text off a page is the cheapest tier. Pulling structure out of it, such as form fields, tables, invoice line items or fields you define yourself, costs more per page. Google Cloud and AWS both publish pay-as-you-go price lists on these terms.

Three other things move the bill. Volume, because some rates step down at high monthly page counts. Custom models, because keeping one deployed can carry an hourly charge on top of the page fee. And region: AWS quotes its Textract pricing examples for US West (Oregon), so check the rate for the region your data actually has to stay in.

The unit is the page, not the document. A three-page invoice whose second and third pages are terms and conditions is billed as three pages, even though only the first holds anything you want.

What does Google charge to host a custom Document AI processor?

Google's Document AI price list includes a hosting rate of $0.05 an hour. Run continuously, that is $438 a year, because a year has 8,760 hours. It is a charge for keeping a processor deployed, and it accrues whether or not any pages go through it.

On its own that is a small number. The shape is the point. Page fees scale with use and fall to zero when nothing is processed. A hosting fee does not. A custom processor deployed for a pilot that ended in the spring is still billing in the autumn unless somebody undeploys it.

Treat it like idle compute. Put it on the monthly cost review, with an owner for each deployed processor and a check on whether anything still calls it.

How do Google Document AI, Amazon Textract and Mistral OCR compare on price?

Google and AWS publish the same basic structure: you pay per page, and the rate rises with the amount of structure you ask for. So compare like with like. Text extraction against text extraction, invoice parsing against invoice parsing, on the same documents.

Google prices Document AI by processor. AWS prices Textract by API, with separate pricing for its expense analysis feature, which is built for invoices and receipts. Mistral stated a list price for OCR 4 in its launch announcement of 23 June 2026. A launch price is a statement made on a date, so confirm it still holds before a budget depends on it.

We have not reproduced per-page rates in this guide. They change, they differ by feature and volume, and a table of them goes stale faster than the advice around it. Google and AWS both include worked examples on their pricing pages. Put your own monthly page count through those, for the feature you will actually use.

What should you check on each vendor's price list?

Why is the API fee usually the smallest part of document AI cost?

Because the model is not the only thing that touches each document. Someone has to check the fields that matter, deal with the documents the model gets wrong, and get the output into the system that needs it. Those are the costs that decide the budget, and none of them appear on a vendor's price list.

Review is the largest. Extraction that is right most of the time still needs a person on the fields where a mistake costs money: totals, bank details, dates, reference numbers. If every document still needs a look, you have not removed the work. You have changed it from typing to checking. That can be worth doing. It is not free.

Exceptions come next. A new supplier layout, a rotated scan, a phone photo of a crumpled receipt, a handwritten amendment: each one falls out of the path the demo showed. How often documents fall out decides how many people you still need, and that matters more than any difference in page rates.

Integration is the third, and it is mostly paid once. Extracted fields are only useful when they land in the accounting system, the case file or the claims queue, validated against records that already exist. That is engineering work, and it is usually larger than the extraction step itself.

What drives the total cost of a document AI system?

• •Page count, including cover sheets, terms and conditions, and appendices nobody reads.
• •The feature you call. Text only is cheapest; forms, tables and custom fields cost more.
• •Layout variety. Ten suppliers with ten formats is a harder problem than one supplier with ten thousand invoices.
• •The accuracy each field needs, which sets how much human review remains.
• •Reprocessing. Sending the same page again is billed again.
• •Hosting fees on custom models left deployed.
• •Integration with the system that uses the output, and monitoring that shows when accuracy drifts.

How do you estimate document AI cost before you commit?

Run a sample of your own documents through the cheapest tier that might work, and measure what comes back. Price lists tell you the page fee. Only your documents tell you the review rate, and the review rate is what decides the budget.

Use a few hundred real documents, not the tidy ones. Include the faxes, the phone photos and the supplier who sends everything sideways. Count pages rather than documents. Run the sample through plain text extraction first, then through the structured tier, and score every field you care about against a copy a person has checked.

Then build the estimate in this order. Monthly pages multiplied by the rate from the vendor's worked example. Plus hosting fees for anything custom. Plus the human time to review fields that came back wrong or low-confidence. Plus the engineering to connect it to the systems downstream. If the last two lines dominate, and they usually do, that is where the design effort belongs.

When is a custom document AI model worth the extra cost?

When pre-built processors miss fields that matter and the volume is high enough to cover the extra work. A custom model costs more than its page rate suggests. Someone has to label training documents and evaluate each version, and on Google's price list a deployed custom processor carries an hourly hosting fee.

Start with the pre-built invoice, receipt or form processors and measure them on your sample. Move to custom only for the fields they consistently get wrong. Then plan for retraining, because documents change. A supplier redesigns its invoice, and the model that was accurate last quarter is not accurate this one.

Is document AI cheaper than processing documents by hand?

Sometimes. It depends on volume, how varied the layouts are and how much review remains, so the honest comparison is against your own current cost per document, measured rather than guessed.

Published benchmarks are a weak substitute. Ardent Partners publishes accounts payable benchmark research, 'AP Metrics That Matter in 2026', but the page presenting it gives no sample size or survey details. That makes it useful for direction and unsuitable as the baseline in a business case.

At low volume it is often not worth doing. If a person handles a few dozen documents a week, build and maintenance will take a long time to pay back, however cheap the pages are. And if the sender can give you structured data, such as an e-invoice, an export or an API, take that instead. Running OCR on a PDF that was generated from a database is paying to reverse a step someone else already did.