Azure Document Intelligence, formerly Form Recognizer, bills per 1,000 pages on the S0 tier: 1.50 dollars for Read OCR, 10 dollars for a prebuilt model such as receipt or invoice, and 30 dollars for a custom model. Every figure on this page was pulled from the Azure retail prices API for East US in August 2026, not from a blog. Upload a receipt below to see the structured output an extraction tool returns without an Azure subscription.
Upload your receipts and invoices
Drop files here or click to upload
Upload failed, please try again
Up to 50 files
Uploading...
Azure does not sell one OCR product at one price. It sells at least six billable models under the Document Intelligence name, each at a different rate, plus commitment tiers, add-on charges, and training time. Most pricing summaries online quote a single number, which is why teams routinely misjudge the bill by a factor of ten.
Read, Layout, prebuilt, custom extraction, custom generative, and the document classifier are priced separately. Reading a page costs 1.50 dollars per 1,000. Running the prebuilt invoice model on that same page costs 10 dollars per 1,000, a 6.7x difference for the same document.
Microsoft renamed Form Recognizer to Document Intelligence, so half the pricing pages, tutorials, and Stack Overflow answers you find still use the old name. The service and the rates are identical, which makes older cost estimates look like a different product.
The F0 free tier analyzes only the first two pages of any document you submit. It is fine for a proof of concept and useless for costing a real workload, so teams that test on F0 often discover the real per page rate only after they switch to S0.
Azure quotes the inference cost. It does not include the developer weeks to wire up the SDK, handle async polling, map the JSON response into your accounting system, or maintain it. For a small finance team that engineering time usually dwarfs the page charges.
These are the S0 pay as you go rates for East US, read directly from the Azure retail prices API in August 2026. Prices are quoted per 1,000 pages. Batch versions of each meter are billed at the same rate as the standard version.
Plain text extraction. Drops to 0.60 dollars per 1,000 pages at higher volume. This is the cheapest meter and it returns text and coordinates only, with no named fields.
The receipt, invoice, ID, W-2, and other prebuilt extractors. This is the meter that matters for receipt and invoice work, and it returns named fields with confidence scores.
Tables, selection marks, and document structure. Same rate as the prebuilt models. Useful when you need the table grid rather than a fixed set of named fields.
Models you train on your own documents. Custom generative extraction is priced identically at 30 dollars per 1,000 pages. Training itself is billed separately at 3 dollars per hour.
Sorts incoming documents by type before extraction. Often overlooked in cost models even though a mixed inbound stream usually needs it.
Extras such as high resolution processing and formula or font detection are billed on top of the base meter. Query fields are charged at 10 dollars per 1,000 pages.
A realistic Azure Document Intelligence estimate takes about five minutes if you count pages rather than documents.
Azure bills per page, not per file. A 3 page invoice is 3 billable pages on the prebuilt meter. Take your monthly document volume and multiply by average page count first.
Tip: Multi page PDFs are where estimates go wrong most often. Check the real average before you budget.
If you need named fields such as vendor, date, tax, and total, you are on the prebuilt meter at 10 dollars per 1,000 pages, not the 1.50 dollar Read meter. Add the classifier if your inbound stream is mixed.
Add the developer time to build and maintain the integration, plus the human minutes spent checking low confidence fields. Microsoft documentation treats a confidence of 0.95 as likely correct 19 times out of 20, so a review step is expected, not optional.
Every rate below comes from the Azure retail prices API for the East US region on the S0 tier. Nothing here is estimated or copied from a third party summary.
You need the real per page meters to decide whether to build on Azure or buy a finished tool. The tables below give you the numbers without wading through the Azure pricing calculator.
You have an Azure bill with Document Intelligence line items and need to know which meter is driving it. Prebuilt at 10 dollars per 1,000 and custom at 30 are the usual culprits.
Above roughly 20,000 pages a month the commitment tiers start to beat pay as you go. The full ladder is below so you can find your break-even point.
If you want receipts and invoices as a spreadsheet and have no interest in running an Azure resource, the per page rate is not the number that decides this. The build effort is.
Azure Document Intelligence bills per page on a meter that depends on which model you call. The table below is the complete S0 pay as you go price list for East US, pulled from the Azure retail prices API in August 2026. Batch versions of each meter carry the same rate as the standard version, so batch processing saves you throughput management, not money.
| Model or meter | Price per 1,000 pages | What it returns |
|---|---|---|
| Read (OCR) | $1.50, then $0.60 at higher volume | Text and coordinates only, no named fields |
| Prebuilt (receipt, invoice, ID, W-2) | $10.00 | Named fields with confidence scores |
| Layout | $10.00 | Tables, selection marks, document structure |
| Custom extraction | $30.00 | Fields you defined and trained |
| Custom generative | $30.00 | Generative field extraction |
| Document classifier | $3.00 | Document type label |
| Query fields | $10.00 | Ad hoc fields requested at call time |
| Add-on features | $6.00 | High resolution, formula and font detection |
| Model training | $3.00 per hour | Billed separately from inference |
The gap that matters most is between Read and prebuilt. Reading the text off a receipt costs 1.50 dollars per 1,000 pages. Getting the vendor, date, tax, and total back as named fields costs 10 dollars per 1,000. If a cost estimate you have been given uses the 1.50 figure for receipt or invoice extraction, it is understating the bill by about 6.7 times. Our OCR API pricing comparison puts these rates side by side with AWS, Google, Veryfi, Mindee, and Nanonets.
Above roughly 20,000 pages a month, commitment tiers beat pay as you go. You buy a monthly block of transactions up front and pay a lower effective rate, with overage billed at that same lower rate. These are the Azure hosted commitment tiers, again from the retail prices API.
| Tier | Monthly block | Price per month | Effective rate per 1,000 |
|---|---|---|---|
| Prebuilt | 20,000 | $190 | $9.50 |
| Prebuilt | 100,000 | $900 | $9.00 |
| Prebuilt | 500,000 | $4,000 | $8.00 |
| Prebuilt | 1,000,000 | $7,500 | $7.50 |
| Read | 500,000 | $375 | $0.75 |
| Read | 2,000,000 | $1,200 | $0.60 |
| Read | 8,000,000 | $4,200 | $0.53 |
| Read | 16,000,000 | $7,200 | $0.45 |
| Custom | 20,000 | $540 | $27.00 |
| Custom | 100,000 | $2,400 | $24.00 |
| Custom | 500,000 | $10,500 | $21.00 |
| Custom | 1,000,000 | $18,000 | $18.00 |
Note how thin the prebuilt discount is. Committing to a million pages a month, a genuinely large workload, moves the rate from 10 dollars to 7.50 dollars per 1,000, a 25% saving. Read commitments are far more aggressive, dropping from 1.50 to 0.45. Azure discounts raw OCR heavily and structured extraction barely, which tells you where the compute cost actually sits.
Take a bookkeeping practice processing 2,500 single page receipts a month. On the prebuilt receipt model that is 2,500 pages at 10 dollars per 1,000, so 25 dollars a month in inference. That number is genuinely low, and it is the honest headline: the Azure API is cheap. What it does not include is an Azure subscription and resource to manage, an SDK integration, async polling for multi page documents, JSON parsing into your accounting format, and someone to maintain all of it. For a two person practice that build is the entire cost. For a platform engineering team processing a million pages, 25 dollars of API cost against a few weeks of build is obviously the right trade.
These are different purchases and it is worth being blunt about which one you need. Azure sells inference: you send a page, you get JSON, and everything either side of that is yours to build. ReceiptOCR sells the finished job: you upload a receipt or invoice in a browser and download a spreadsheet. Per page, Azure is cheaper. Per project, it depends entirely on whether you have engineers.
| Azure Document Intelligence | ReceiptOCR | |
|---|---|---|
| What you get back | JSON with fields and confidence scores | Excel, CSV, or a QuickBooks ready file |
| Setup required | Azure subscription, resource, SDK integration | Upload in a browser, no account setup to process |
| Per page cost | $0.01 on the prebuilt meter | Included in a flat monthly plan |
| Pricing model | Pay as you go or monthly commitment | Flat monthly plan with a page allowance |
| Best for | Engineering teams building a pipeline | Finance teams who need the spreadsheet |
| Scales to millions of pages | Yes, with commitment tier discounts | Built for business volumes, not web scale |
If you are comparing Azure against the other cloud vendors rather than against a finished tool, the honest per page comparison lives on our OCR API pricing page, and the reasons vendor accuracy claims are hard to compare are covered under OCR accuracy. If you have already decided Azure is more machinery than you need, the Azure Document Intelligence alternative page covers the switch.
Azure wins on breadth and on scale. It handles far more document types than a receipt and invoice tool: IDs, W-2s, health insurance cards, contracts, and anything you care to train a custom model on. It runs inside your own Azure tenant, which matters for procurement and data residency in regulated industries. It offers disconnected containers for air gapped environments, priced annually rather than per page. And at very high volume the commitment tiers take Read down to 0.45 dollars per 1,000 pages, which no finished tool will match. If you have engineers, high volume, and varied document types, build on Azure.
On the S0 pay as you go tier in East US, Azure Document Intelligence costs 1.50 dollars per 1,000 pages for Read OCR, 10 dollars per 1,000 pages for a prebuilt model such as receipt or invoice, and 30 dollars per 1,000 pages for a custom model. The document classifier is 3 dollars per 1,000 pages and model training is 3 dollars per hour.
There is no difference. Microsoft renamed Form Recognizer to Azure AI Document Intelligence, and it is the same service at the same rates. Older documentation and pricing summaries still say Form Recognizer, which is why the two names appear side by side, but a quote for one applies to the other.
Read OCR costs 0.0015 dollars per page, or 1.50 dollars per 1,000 pages, dropping to 0.0006 dollars per page at higher volume. Structured extraction with a prebuilt receipt or invoice model costs 0.01 dollars per page, which is 10 dollars per 1,000. The gap between plain text and named fields is roughly 6.7 times.
There is an F0 tier at no cost, but it analyzes only the first two pages of any document you submit. That makes it usable for a proof of concept and unsuitable for costing or running a real workload. Any production volume runs on the S0 tier and is billed per page at the rates above.
Per page. A three page invoice submitted to the prebuilt invoice model is billed as three pages, not one document. This is the single most common source of underestimated Azure OCR bills, so multiply your monthly document count by average page count before budgeting.
Commitment tiers start to pay above roughly 20,000 pages a month. The entry prebuilt tier is 190 dollars for 20,000 transactions, an effective 9.50 dollars per 1,000 against the 10 dollar pay as you go rate. Discounts deepen with volume, reaching 7.50 dollars per 1,000 at a million pages a month.
Use the prebuilt receipt or prebuilt invoice model, billed at 10 dollars per 1,000 pages. The Read meter at 1.50 dollars returns text and coordinates but no named fields, so you would have to write your own parsing to find the vendor, date, tax, and total. Custom models at 30 dollars are only worth it for document layouts the prebuilt models do not cover.
No. Azure publishes an estimated accuracy score only for custom template models, calculated against your own training data. Custom neural and generative models return no accuracy score at all, and prebuilt models return per field confidence values. Microsoft documentation treats a confidence of 0.95 as likely correct 19 times out of 20, so budget for a human review step.
Stop typing receipts by hand
Upload your receipts and invoices and get a clean Excel or CSV file in minutes.
Extract my receipts nowFree to try, no sign up required
Azure, AWS, Google, Veryfi and Mindee rates compared per page.
Receipt and invoice data to Excel without an Azure resource.
How the four product categories sold as OCR actually differ.
Why no major vendor publishes a receipt extraction accuracy rate.
Extract invoice fields without building on a cloud vendor SDK.
