Document AI Compared: Textract, Google and Azure

For a one-page invoice, Amazon Textract and Azure AI Document Intelligence both charge one cent. Google Document AI charges ten cents, because Google's billing unit is a document of up to ten pages, not a page. At ten pages per document all three cost exactly the same. Almost every real invoice is one or two pages, so that difference decides the comparison before anything about accuracy comes up.

Every price below was read from the vendor's own page or price API in September 2026, and every link goes to the source. One of the three does not publish a price you can read without picking a region first; that is covered below.

The three published rates

Amazon Textract's pricing page, September 2026, shows the Analyze Expense API — the one built for invoices and receipts — at $0.01 per page for the first 1 million pages a month and $0.008 per page above that. The plain Detect Document Text API, which returns raw text and no structured fields, is $0.0015 per page and then $0.0006. Analyze ID is $25 per 1,000 pages for the first 100,000. The page presents its figures for the US West (Oregon) region.

Google Document AI's pricing page, September 2026, lists the Invoice parser and the Expense parser — the latter is the former receipt parser — at $0.10 per count, with the footnote "1 count equals up to 10 pages in a document." The page's own worked example is explicit: "To parse a document with between 1 to 10 pages costs $0.10." Enterprise Document OCR, the text-only processor, is free for the first 1,000 counts a month, then $1.50 per 1,000 and $0.60 above 5 million. Form Parser and Custom Extractor are $30 per 1,000 pages up to 1 million.

Azure is the awkward one. The Azure AI Document Intelligence pricing page renders every figure as "$-" until you choose a region, so a plain fetch of the URL returns no prices at all. Selecting a region in a browser fills the table in. Read that way on 4 September 2026, with region East US and currency USD, the free F0 tier and the paid S0 tier price as follows. Add-on features and query fields are separate meters that bill on top of whichever model you run, not alternatives to it.

The prebuilt invoice and receipt models are the prebuilt meter, so Azure matches Textract at a cent a page. Custom extraction at $30 per 1,000 also matches Google's Custom Extractor exactly.

One inconsistency to know about before you quote an Azure number. Microsoft also publishes prices through its Retail Prices API, still filed under the service's former name, Form Recognizer. Queried the same month for East US, the API returns "S0 Pre-built Pages" at $10.00 per 1,000 — agreeing with the page — but "S0 Custom Pages" at $50.00 per 1,000, where the page says $30. We have not found a Microsoft statement reconciling the two, and the likeliest explanation is that the API still carries a legacy Form Recognizer meter. The pricing page is the customer-facing document and is the figure quoted above; if you are budgeting custom extraction on Azure at volume, that $20 per 1,000 gap is worth confirming with Microsoft rather than inferring from either source. The API also shows US Government regions billed higher, at $12.50 and $62.50 per 1,000.

What that costs at real volumes

Take 500 pages a month, then 5,000 — a small firm's accounts payable pile and a busy bookkeeping practice. Because Google's unit is a document, the answer depends on how long your documents are, so here are both ends of it. This is arithmetic from the published rates, not a benchmark.

The pattern is the whole story. Google is ten times more expensive on single-page documents, five times on two-page documents, and exactly level at ten. If you process contracts, statements or long remittance packs, Google's unit works in your favor. If you process the ordinary one-page supplier invoice that most small businesses actually receive, it does not.

The free tiers are not comparable

All three advertise something free. None of them mean the same thing.

Textract offers a free tier that, in the page's wording, "lasts for three months," for new AWS customers only: 1,000 pages a month of Detect Document Text, 100 pages a month of Analyze Expense, and 100 pages a month of Analyze ID. One hundred invoices a month for three months is a pilot, not a running service.

Azure is the only one with a standing free allowance. Its pricing page lists a Free (F0) tier at "0 - 500 pages free per month," with no expiry. The catch is in Microsoft's service limits: F0 caps documents at 2 pages and 4 MB and throttles you to 1 transaction per second. For a business receiving under 500 pages a month of one- and two-page invoices, F0 is a genuinely free production tier. Anything longer than two pages fails on the free tier and needs S0.

Google publishes no free monthly allowance for the Invoice or Expense parser. The free element is the general Google Cloud free-trial credit for new customers (Google Cloud free program), which is a one-off trial balance, not a recurring allowance. Enterprise Document OCR does get 1,000 free counts a month, but that processor returns text, not invoice fields.

Page and file limits worth checking first

These caps disqualify architectures more often than price does.

Textract's fixed quotas allow just 1 page per document for synchronous calls, and up to 3,000 pages asynchronously. PDFs and TIFFs may be 10 MB synchronous, 500 MB asynchronous. If you want an immediate answer, you are sending single pages.

Google's documented limits put the Invoice parser at a maximum of 15 pages for online requests and 200 for batch, and the Expense parser at 10 pages either way. File size is 40 MB online and 1 GB for batch.

Azure S0 allows up to 2,000 pages per document and 500 MB, at a default 15 analyze transactions per second that can be raised on request.

What they extract

All three return structured invoice fields rather than raw text, and the field lists overlap heavily. Amazon's AnalyzeExpense documentation names the specific types it returns, including VENDOR_NAME, INVOICE_RECEIPT_ID, INVOICE_RECEIPT_DATE, DUE_DATE, PO_NUMBER, SUBTOTAL, TAX, TOTAL, AMOUNT_DUE, PAYMENT_TERMS, and line-item types such as ITEM, QUANTITY, UNIT_PRICE and PRODUCT_CODE.

Google's Invoice parser and Azure's prebuilt invoice model publish comparable schemas covering vendor, dates, totals, tax and line items. What none of the three publishes is a per-field accuracy figure you could use to rank them on your documents. There is no vendor-published benchmark that settles it, and any article claiming one tool is more accurate than another without naming a dataset is guessing. The only test that means anything is fifty of your own invoices through each API.

Where it is a tie

At ten pages per document, all three cost the same to the cent. Below roughly 500 pages a month the whole comparison is worth under $50 either way, which is less than an hour of the time you would spend choosing. If you are already on AWS, Google Cloud or Azure, staying inside that account is worth more than the price gap, because it removes a second vendor, a second bill and a second security review.

It is also a tie on the thing people expect to be decisive. All three are per-use APIs with no minimum commitment, no seat licences and no annual contract at these tiers. You can put a hundred invoices through each and pay a few dollars to find out.

What these prices do not include

Extraction is one line on the bill. Storage of the source PDFs, data egress, the queue or function that calls the API, and whatever automation platform moves the extracted fields into your accounting system are all extra, and on small volumes the platform usually costs more than the OCR. Our walkthrough of how to automate invoice data entry covers the full pipeline, including the exception queue that decides whether the project works. The pattern of underestimating everything around the headline price is the subject of the hidden costs of automation platforms.

One last practical note: re-check the numbers yourself. Textract's page shows Oregon pricing, Azure's shows nothing until you pick a region, and both vary by region. All figures here carry a September 2026 date.