Azure and AWS both charge $10.00 per 1,000 pages for their invoice models. Google charges $0.10 per document in 10-page blocks, which is $100.00 per 1,000 one-page invoices. Every rate here was read from the vendors' own pricing sources on 2 September 2026, converted to cost per 1,000 invoices, and paired with independent line-item accuracy. Drop an invoice below to see the output first.
Upload your invoices
Drop files here or click to upload
Up to 50 files
Uploading...
Google, Microsoft and Amazon all sell an invoice extraction model, and all three publish a price. The prices are quoted in different units, on pages that in one case render no numbers at all, and the discounts that matter most are not on the pricing pages. Here is what trips people up.
Google quotes per document, Azure and AWS quote per 1,000 pages, and Azure's own pricing page renders no numbers at all. You cannot line the three up without converting them to the same unit and modeling your real page distribution.
AWS Textract publishes a $1.50 OCR rate that has nothing to do with invoices, and a $50.00 Forms rate that returns less invoice structure than the $10.00 Analyze Expense meter. Picking the wrong meter on the same service can multiply your bill by seven.
Azure's commitment tiers, which beat pay-as-you-go by 24% from 20,000 pages a month, appear in the Retail Prices API and nowhere on the pricing page. Most comparisons never mention they exist.
Vendors and roundups quote a single accuracy percentage. Invoice extraction fails on line items, not on the vendor name, and the split between the two is where the services actually differ.
One unit, real numbers, and the parts of each price list the vendors bury. Nothing here is estimated or repeated from another comparison page.
Every figure on this page was read from cloud.google.com/document-ai/pricing, the Azure Retail Prices API and the AWS Price List API, then converted to cost per 1,000 invoices at several document lengths so the three are directly comparable.
Google's 10-page block is modeled at 1, 2, 5, 10, 11, 15, 20 and 21 pages per invoice, which shows exactly where it ties the per-page vendors and where it costs ten times as much.
Independent accuracy is reported as two numbers, because a service that reads the vendor name perfectly and drops half the line items is not accurate for accounts payable.
The Azure commitment ladder and disconnected-container pricing are laid out in full, including the finding that Microsoft charges no premium at all for running air-gapped.
The right answer depends on your document mix, so model it rather than trusting a headline rate.
Pull a month of vendor bills and record how many pages each one has. Azure and AWS bill every page. Google bills each document as a block of up to ten.
Pages times $0.010 gives you Azure and AWS. Documents times $0.10, rounding each document up to the next ten pages, gives you Google.
Tip: If most of your invoices are one or two pages, expect Google to land near ten times the other two.
None of the three prices includes authentication, entity mapping, line-item flattening, error handling or export. Price that engineering time, then compare the total against a tool that already returns a spreadsheet.
Anyone about to commit budget or engineering time to one of the three cloud invoice parsers.
You need the true unit cost before you commit to a cloud, including the hosting rent Google charges for a deployed custom processor and the reserved-capacity tier.
Your invoices are one or two pages, which is the worst case for block billing. The first table tells you what each service costs on your actual document mix.
If outbound network access is off the table, only one of these three can run disconnected, and it costs the same as the connected version.
Commitment tiers, volume breaks and where they kick in, so you know whether a contract is worth negotiating at your volume.
For a normal one-page or two-page US vendor invoice, Azure AI Document Intelligence and AWS Textract both cost $10.00 per 1,000 pages, while Google Document AI charges $0.10 per document, which works out to $100.00 per 1,000 invoices. On short invoices Google is ten times the price of the other two. The gap closes only when your documents actually run ten pages long, because Google bills in 10-page blocks rather than per page. Last updated September 2026. Every rate below was read from cloud.google.com/document-ai/pricing, the Azure Retail Prices API and the AWS Price List API on 2 September 2026.
This is the table that decides the choice, and it is the one nobody publishes, because you have to model your own page distribution to build it. Azure and AWS bill per page. Google bills per document, rounding up to the next block of ten pages.
| Pages per invoice | Google Document AI (Invoice Parser) | Azure AI Document Intelligence (prebuilt invoice) | AWS Textract (Analyze Expense) | Cheapest |
|---|---|---|---|---|
| 1 page | $100.00 | $10.00 | $10.00 | Azure and AWS, by 10x |
| 2 pages | $100.00 | $20.00 | $20.00 | Azure and AWS, by 5x |
| 5 pages | $100.00 | $50.00 | $50.00 | Azure and AWS, by 2x |
| 10 pages | $100.00 | $100.00 | $100.00 | Three-way tie |
| 11 pages | $200.00 | $110.00 | $110.00 | Azure and AWS |
| 15 pages | $200.00 | $150.00 | $150.00 | Azure and AWS |
| 20 pages | $200.00 | $200.00 | $200.00 | Three-way tie |
| 21 pages | $300.00 | $210.00 | $210.00 | Azure and AWS |
Read the right-hand column again. Google's block model ties at exact multiples of ten pages and loses at every other length. It is never cheaper than per-page billing on the same document. That is a structural property of rounding up, not a quirk of one price point, and it does not change at any volume because the Invoice Parser has no volume discount at all.
US accounts payable is dominated by one-page and two-page invoices, which is exactly where the model is worst. If you process 5,000 single-page vendor invoices a month, Google charges $500 and the other two charge $50.
Google prices its pretrained processors per count, and its own pricing page defines the unit precisely: "1 count equals up to 10 pages in a document." The worked examples Google publishes are unambiguous. A document of 1 to 10 pages costs $0.10. A document of 11 to 20 pages costs $0.20. A document of 91 to 100 pages costs $1.00.
Two practical consequences follow. First, a single-page invoice consumes a whole block, so you pay for ten pages and use one. Second, batching does not help, because the block is scoped to a document, not to a request. Merging fifty single-page invoices into one PDF would cut the count, but then the parser returns one merged result and you lose the per-invoice separation you needed in the first place.
There is also a hard limit worth knowing before you design around it. Documents longer than ten pages are not supported by synchronous requests at all. You have to use a batch request, which caps out at 200 pages per document.
All figures in US dollars per 1,000 pages unless the row says otherwise. Google's pretrained rows are per document of up to ten pages, which is why they are marked separately.
| What you are buying | Google Document AI | Azure AI Document Intelligence (S0) | AWS Textract |
|---|---|---|---|
| Plain text OCR | $1.50 (Enterprise Document OCR, first 5M), $0.60 above 5M | $1.50 (Read), $0.60 above 1M | $1.50 (Detect Document Text), $0.60 above 1M |
| Invoice model with line items | $0.10 per document up to 10 pages (Invoice Parser) | $10.00 (prebuilt invoice) | $10.00 (Analyze Expense), $8.00 above 1M |
| Receipts and expenses | $0.10 per document (Expense Parser) | $10.00 (prebuilt receipt) | $10.00 (Analyze Expense) |
| Utility bills | $0.10 per document (Utility Parser) | Not a prebuilt model | Not a prebuilt model |
| Bank statements | $0.75 per classified document | Not a prebuilt model | $70.00 (Analyze Lending) |
| Generic layout and structure | $10.00 (Layout Parser, chunking included) | $10.00 (Layout) | $4.00 (Layout) |
| Generic forms and key-value pairs | $30.00 (Form Parser), $20.00 above 1M | $10.00 (query fields) | $50.00 (Forms), $40.00 above 1M |
| Tables | Included in Form Parser | Included in Layout | $15.00 (Tables), $10.00 above 1M |
| Your own trained schema | $30.00 (Custom Extractor), $20.00 above 1M | $30.00 (custom extraction) | $25.00 (Custom Queries), $15.00 above 1M |
| Document classification | $5.00 (custom classifier), $3.00 above 1M | $3.00 (document classifier) | Not a separate meter |
| Training | Included; you pay hosting instead | $3.00 per hour | Included |
| Hosting a custom model | $0.05 per hour, about $438 per year per deployed version | None | None |
| Reserved capacity | $0.410958904 per hour (provisioned tier) | Commitment tiers, see below | None |
Two rows deserve a second look. Google's Form Parser at $30.00 per 1,000 pages is three times the Azure prebuilt invoice model, and it returns generic key-value pairs rather than mapped invoice fields, so it is the wrong tool for AP even though people reach for it. And Google is the only one of the three that charges rent: $0.05 an hour to keep a custom processor version deployed, whether or not you send it a single page.
Price only matters if the output is usable. The most useful independent test we have found on scanned invoices is BusinesswareTech's January 2025 benchmark, which scored the three cloud services on header fields and on line items separately. That split matters more than the headline number, because header fields are easy and line items are where invoice extraction actually fails.
| Service | Header field accuracy | Line-item accuracy |
|---|---|---|
| Azure AI Document Intelligence | 93.0% | 87.0% |
| AWS Textract | 78.0% | 82.0% |
| Google Document AI | 82.0% | 40.0% |
| GPT-4o with third-party OCR | 98.0% | 57.0% |
Google Document AI came last on line items by a wide margin, recovering 40.0% where Azure recovered 87.0%. Put that next to the pricing table and the conclusion is uncomfortable for Google on this particular job: on a one-page invoice it is ten times the price of Azure and less than half as good at the part of the document that is hardest to read.
One caveat, stated plainly because it matters. That benchmark is a single third-party study on scanned documents from January 2025, and all three vendors have shipped model updates since. Treat it as a strong directional signal, not as a warranty, and run your own sample before you commit. Nobody, including us, can tell you how a given service will do on your vendors' layouts without testing them.
This is where the three diverge most, and where the published pages are least helpful.
Google Document AI has no volume discount on the Invoice Parser. It is $0.10 per document at ten invoices a month and at ten million. The volume breaks Google does publish apply to other processors: Enterprise Document OCR drops from $1.50 to $0.60 above five million pages, and Custom Extractor and Form Parser drop from $30.00 to $20.00 above one million.
AWS Textract discounts Analyze Expense from $10.00 to $8.00 above one million pages a month, applied automatically with no contract to sign.
Azure is the only one of the three that sells commitment tiers, and Microsoft does not show them on the pricing page. They are visible in the Retail Prices API. For the prebuilt models, which is the invoice meter:
| Commitment | Azure-hosted | Effective per 1,000 pages | Connected container | Effective per 1,000 pages |
|---|---|---|---|---|
| 20,000 pages/month | $190/mo | $9.50 | $152/mo | $7.60 |
| 100,000 pages/month | $900/mo | $9.00 | $720/mo | $7.20 |
| 500,000 pages/month | $4,000/mo | $8.00 | $3,200/mo | $6.40 |
| 1,000,000 pages/month | $7,500/mo | $7.50 | $6,000/mo | $6.00 |
A 20,000-page commitment in a container already beats pay-as-you-go by 24%, and 20,000 pages a month is an ordinary mid-sized AP department, not an enterprise. Overage is billed at that tier's effective rate, so exceeding a commitment is not punished.
Only Azure can run disconnected. Microsoft publishes annual disconnected (air-gapped) prices for the same containers, and when you annualize them they come out identical to the connected monthly rate: $8,640 a year for the 100,000-page tier is exactly twelve times $720. Microsoft charges no premium for running Document Intelligence with no outbound network access. Google Document AI and AWS Textract are cloud-only, so if your auditors require an air-gapped pipeline, the comparison ends here and Azure wins by default.
| Service | Free allowance on the invoice model | Free allowance elsewhere | Expires? |
|---|---|---|---|
| Google Document AI | None. The Invoice Parser bills from the first document. | First 1,000 counts of Enterprise Document OCR are $0.00 | No, the OCR allowance is ongoing |
| Azure AI Document Intelligence | None on S0. A separate Free (F0) SKU exists at $0.00 but reads only the first two pages of a document. | Same F0 constraint applies | No |
| AWS Textract | 100 pages per month of Analyze Expense | 1,000 pages/month Detect Document Text, 2,000 pages/month Analyze Lending, 100 pages/month Forms and Tables | Yes, three months for new AWS accounts |
AWS has the only genuinely useful free trial of an invoice model, and it is small and time-boxed. Azure's F0 two-page cap makes it useless for testing anything but a one-page invoice. Google gives you nothing on the invoice path.
Choose Azure AI Document Intelligence if invoices are the job. It had the best line-item accuracy in the independent test, it matches AWS on the headline rate, it is the only one with commitment pricing that a mid-sized AP department can reach, and it is the only one that runs air-gapped.
Choose AWS Textract if you are already deep in AWS and want the lowest-friction path, or if you clear a million pages a month, where Analyze Expense drops to $8.00 with no negotiation. Be aware of the trap in its own price list: running Forms plus Tables on an invoice costs $70.00 per 1,000 pages and returns less invoice structure than Analyze Expense does at $10.00.
Choose Google Document AI when your documents are genuinely long, when you are standardizing on Google Cloud for reasons that outrank per-invoice cost, or when you need one of the specialist parsers the other two do not offer, such as the Utility Parser or the Bank Statement Parser. For short vendor invoices it is the hardest of the three to justify on price.
Choose none of the three if you do not have engineers to spend on this. All three prices above are for a raw API. None of them includes the work of authenticating, mapping entities to your chart of accounts, flattening line items, handling failures, or producing a spreadsheet an accountant can use. For most finance teams that build-and-maintain cost is larger than the processing charge it was meant to save, which is the reason a finished extraction tool exists at all. You can upload an invoice at the top of this page and see the structured output before you write any code.
No, not for invoices. The pretrained Invoice Parser bills $0.10 per document from the very first file, with no free allowance and no volume discount. Google does give you the first 1,000 counts of the Enterprise Document OCR processor at $0.00, but that processor returns raw text, not mapped invoice fields like vendor, invoice number, dates, tax and line items.
$0.10 per invoice for documents up to ten pages, because Google bills its pretrained processors per count and one count equals up to ten pages in a document. A one-page invoice and a ten-page invoice both cost $0.10. An eleven-page invoice costs $0.20. That works out to $100.00 per 1,000 typical single-page vendor invoices.
For invoice extraction, Azure AI Document Intelligence is the stronger choice on both price and accuracy. It costs $10.00 per 1,000 pages against Google's effective $100.00 per 1,000 single-page invoices, and in BusinesswareTech's January 2025 benchmark on scanned invoices it recovered 87.0% of line items where Google recovered 40.0%. Google is more competitive on long documents and on specialist parsers Azure does not offer.
On short invoices, yes, by ten times. AWS Textract Analyze Expense is $10.00 per 1,000 pages, so 1,000 single-page invoices cost $10.00. Google charges $0.10 per document regardless of length, so the same 1,000 invoices cost $100.00. The two are level only when your documents run exactly ten or twenty pages, and Textract also discounts to $8.00 above one million pages, which Google never does.
Yes. The Invoice Parser returns line-item entities alongside header fields such as supplier name, invoice ID, dates, subtotal, tax and total. Accuracy on those line items is the weak point: the January 2025 BusinesswareTech benchmark on scanned invoices scored Google at 40.0% on line items, the lowest of the cloud services tested, against 87.0% for Azure and 82.0% for AWS Textract.
Cloud Vision does general image analysis and returns raw detected text with no document understanding. Document AI is the document-specific platform, and its processors return structured entities. For invoices the relevant one is the Invoice Parser, which maps fields to a defined invoice schema. Running Cloud Vision on an invoice gives you a wall of text you would still have to parse yourself.
There is a Free (F0) SKU priced at $0.00, but it reads only the first two pages of any document, so it is only useful for smoke-testing single-page files. The paid S0 tier, which is where the $10.00 per 1,000 pages prebuilt invoice model lives, has no free allowance. AWS Textract is the only one of the three with a usable free trial of an invoice model, at 100 Analyze Expense pages a month for three months on new accounts.
No. Google Document AI and AWS Textract are cloud-only services. Azure AI Document Intelligence is the only one of the three that ships containers you can run connected or fully disconnected, and Microsoft charges no premium for the air-gapped option: the annual disconnected price divided by twelve equals the connected monthly price exactly. If an auditor requires no outbound network access, Azure is the only option among the three.
Google prices its pretrained processors per count, and defines one count as up to ten pages in a document. The model suits long documents such as contracts and loan files, where a ten-page block is filled efficiently. It works against you on accounts payable, where most vendor invoices are one or two pages and each one still consumes a full block.
AWS Textract drops Analyze Expense from $10.00 to $8.00 per 1,000 pages above one million pages a month, applied automatically. Azure sells commitment tiers that are not shown on its pricing page, starting at $190 a month for 20,000 pages, an effective $9.50, and reaching $7.50 per 1,000 at one million pages a month, with cheaper container rates. Google Document AI's Invoice Parser has no volume discount at any level.
Every Document AI processor rate, and why the 10-page block model rounds against short invoices.
The full Azure rate card from the Retail Prices API, including tiers the pricing page hides.
All sixteen Textract feature rates, including the Forms plus Tables trap on invoices.
The other cloud head to head, with API limits and page caps compared.
Skip the Google Cloud build and get mapped invoice fields in a spreadsheet instead.
An API that already returns mapped invoice fields, with no entity schema to build.
Start turning your invoices into clean, structured spreadsheet data.
USD
per month
per month
$288 charged today
Choose speed vs accuracy when extracting
| Base AI Faster | 2,500 pages |
| Pro AI Best accuracy | 500 pages |
Scale invoice extraction across your whole team with automation.
USD
per month
per month
$888 charged today
Choose speed vs accuracy when extracting
| Base AI Faster | 10,000 pages |
| Pro AI Best accuracy | 2,000 pages |
Enterprise‑grade invoice extraction, security, and controls.
USD
per month
per month
$ charged today
Choose speed vs accuracy when extracting
| Base AI Faster | pages |
| Pro AI Best accuracy | pages |
Secure checkout powered by Stripe
