Form Recognizer became Azure AI Document Intelligence in July 2023, and the rates did not change: $1.50 per 1,000 pages for OCR, $10.00 for layout and prebuilt models, $30.00 for custom extraction. Microsoft's billing feed still carries the old product name too, at a different rate. Every number below came from that feed, not from a marketing page.
Upload a document to extract
Drop files here or click to upload
Up to 50 files
Free plan extracts the first 5, rest can be unlocked after
Uploading...
A per-page rate tells you nothing about whether the fields come out right on your paperwork. Drop a document in and read the output before you commit to a vendor.
Azure Form Recognizer is now Azure AI Document Intelligence, and the rename did not change what you pay. On the Standard (S0) tier, plain OCR with the Read model costs $1.50 per 1,000 pages, falling to $0.60 above a million pages a month. Layout analysis with tables and structure costs $10.00 per 1,000 pages, and so does any prebuilt model such as invoice, receipt, ID or W-2. A custom model you trained yourself costs $30.00 per 1,000 pages, document classification costs $3.00, and each add-on capability adds $6.00. There is a Free (F0) tier, but it returns only the first two pages of any request, so it is for testing the output shape rather than running work.
The number that surprises people is on the old product name. Microsoft's price feed still lists a live Form Recognizer product whose custom-pages meter bills $50.00 per 1,000, twenty dollars above the $30.00 charged under the Document Intelligence name for the same job.
Rates read from Microsoft's Azure Retail Prices API on 10 August 2026, region East US.
Yes. Microsoft renamed Azure Form Recognizer to Azure AI Document Intelligence in July 2023, and the documentation now files it under Azure Document Intelligence in Foundry Tools. Same endpoints, same models, same rates. The reason the old name refuses to die is that Microsoft never removed it from the places engineers and finance teams actually look.
| Where you see it | The name it uses today | Note |
|---|---|---|
| Product documentation | Azure Document Intelligence in Foundry Tools | Learn article titles read "(formerly Azure AI Form Recognizer)" |
| Billing feed serviceName | Foundry Tools | Both product names now bill under the Foundry Tools service |
| Billing feed productName | Azure Document Intelligence, and Form Recognizer | Two separate live products, and they do not share every rate |
| .NET SDK package | Azure.AI.FormRecognizer 4.1.0 | The old name, still current on NuGet |
| Python SDK package | azure-ai-formrecognizer 3.3.0 | The old name, still current on PyPI |
| JavaScript SDK package | @azure/ai-form-recognizer 5.0.0 | The old name, still current on npm |
| v2.x client classes | FormRecognizerClient, FormTrainingClient | v3.0 and later renamed these to DocumentAnalysisClient |
| Support tag Microsoft asks you to use | azure-form-recognizer | Learn: "tag it with azure-form-recognizer" |
| Pricing URL linked from Learn | /pricing/details/form-recognizer/ | The service-limits page still links the old pricing path |
If your code still imports FormRecognizerClient or FormTrainingClient, you are on a v2.x client. Version 3.0 replaced both with DocumentAnalysisClient, and REST API v2.1 reaches end of support on September 15, 2027.
Because Microsoft never retired the old billing product. Query the Azure Retail Prices API today and two separate products come back, both under the service name Foundry Tools: Azure Document Intelligence with 68 meters, and Form Recognizer with four. They do not charge the same rate for custom extraction.
| Meter | Under "Form Recognizer" | Under "Azure Document Intelligence" |
|---|---|---|
| S0 Custom Pages | $50.00 per 1,000 | $30.00 per 1,000 |
| S0 Pre-built Pages | $10.00 per 1,000 | $10.00 per 1,000 |
| Meter effective from | 1 August 2021 | 1 June 2024 |
| Number of live meters | 4 | 68 |
It proves the two meters exist, at two rates, in Microsoft's own billing feed. We re-ran the query in East US, West US 2 and North Europe and got the same answer in all three.
It does not prove which meter your subscription is billed on. That depends on how your resource was provisioned. So do not assume you are overpaying, and do not assume you are not: open your Azure invoice, find the line item, and read the meter name rather than the product heading.
At 100,000 custom pages a month the gap between the two is $2,000 a month.
Microsoft's marketing pricing page is a JavaScript application that renders its rates client side, which makes it a poor citation. The Azure Retail Prices API is the feed the portal bills from, and it is public. These are the Standard (S0) pay-as-you-go rates it returns for East US.
| What you are doing | Meter name on your bill | Per 1,000 pages | Volume behavior |
|---|---|---|---|
| Read (plain OCR, text and layout-free) | S0 Read Pages | $1.50 | Drops to $0.60 above 1,000,000 pages a month |
| Layout (tables, structure, reading order) | S0 Layout Pages | $10.00 | Single tier, no volume break |
| Prebuilt models (invoice, receipt, ID, W-2, bank statement) | S0 Pre-built Pages | $10.00 | Single tier |
| Custom extraction (your own trained model) | S0 Custom Pages | $30.00 | Single tier. Commitment tiers go lower |
| Custom generative extraction | S0 Custom Generative Pages | $30.00 | Single tier |
| Document classifier | S0 pages for doc classifier | $3.00 | Cheapest classification rate of the big three |
| Query fields (ask for a field by name) | S0 pages for query fields | $10.00 | See the $200 legacy meter warning below |
| Add-on capabilities (each) | S0 Add-on for Pages | $6.00 | Barcodes, formulas, fonts, high resolution, language |
| Custom neural training | S0 Training | $3.00 per hour | First 10 hours each month are free on v4.0 |
One correction worth making, because it circulates widely: custom extraction has no volume tier. The feed returns a single S0 Custom Pages row at $30.00 with a tier minimum of zero. Read is the only meter with a genuine pay-as-you-go volume break, and it takes a million pages a month to reach it. Everything below $30.00 for custom work comes from a commitment tier instead.
No. Azure publishes a separate batch meter for every operation, and each one is priced identically to its synchronous twin. The Batch API is a throughput and convenience feature, not a pricing feature. This catches out teams migrating from an LLM provider, where batch is the standard way to halve a bill.
| Operation | Synchronous | Batch | Saving |
|---|---|---|---|
| Read | $1.50 | $1.50 | No discount |
| Layout | $10.00 | $10.00 | No discount |
| Prebuilt | $10.00 | $10.00 | No discount |
| Custom extraction | $30.00 | $30.00 | No discount |
| Custom generative | $30.00 | $30.00 | No discount |
| Document classifier | $3.00 | $3.00 | No discount |
| Query fields | $10.00 | $10.00 | No discount |
| Add-on capabilities | $6.00 | $6.00 | No discount |
Mistral, Anthropic and OpenAI all price batch work at roughly half the synchronous rate. AWS Textract and Google Document AI, like Azure, do not discount it either. So batch pricing is an LLM-provider convention rather than an OCR-vendor one, and moving a workload to Azure's Batch API to save money will not save any.
The one thing batch does change is the ceiling: it lets you hand Azure a set of documents instead of managing 15 transactions a second yourself.
Yes, and it crosses over earlier than most buyers assume. Read on pay as you go is $1.50 per 1,000 pages and only drops to $0.60 above a million pages a month. The Read 500K commitment is $375 a month for 500,000 pages, which is $0.75 per 1,000. Since 250,000 pages of pay-as-you-go Read also costs $375, the commitment doubles your pages for the same money at that point.
| Model | Tier | Azure, per month | Effective per 1,000 | Connected (container) | Effective per 1,000 |
|---|---|---|---|---|---|
| Read | 500K pages | $375 | $0.75 | $338 | $0.68 |
| Read | 2,000K pages | $1,200 | $0.60 | $1,080 | $0.54 |
| Read | 8,000K pages | $4,200 | $0.53 | $3,780 | $0.47 |
| Read | 16,000K pages | $7,200 | $0.45 | $6,480 | $0.41 |
| Prebuilt | 20K pages | $190 | $9.50 | $152 | $7.60 |
| Prebuilt | 100K pages | $900 | $9.00 | $720 | $7.20 |
| Prebuilt | 500K pages | $4,000 | $8.00 | $3,200 | $6.40 |
| Prebuilt | 1,000K pages | $7,500 | $7.50 | $6,000 | $6.00 |
| Custom | 20K pages | $540 | $27.00 | $459 | $22.95 |
| Custom | 100K pages | $2,400 | $24.00 | $2,040 | $20.40 |
| Custom | 500K pages | $10,500 | $21.00 | $8,925 | $17.85 |
| Custom | 1,000K pages | $18,000 | $18.00 | $15,300 | $15.30 |
Two things to know before quoting these. Microsoft's documentation says container pricing matches cloud service pricing, and on pay-as-you-go meters it does. The commitment SKUs are the exception: the Connected (container) units run 15 to 20% below the Azure ones across every model. Second, going over your committed volume does not stop the meter, it bills an overage rate that is close to, but not the same as, the tier's effective rate.
There is a Free (F0) tier, and it is a testing tier rather than a free allowance. Microsoft states it plainly: the free tier limits the analyze response to only the first two pages in a request. Send a ten-page contract and you get two pages back. It is genuinely useful for checking the shape of the JSON before you write code against it, and useless for processing anything real.
| Quota | Free (F0) | Standard (S0) |
|---|---|---|
| Pages returned per analyze request | First 2 pages only | 2,000 |
| Max document size | 4 MB | 500 MB |
| Analyze transactions per second | 1 | 15 by default, adjustable |
| Get operations per second | 1 | 50 by default, adjustable |
| Max template models | 500 | 5,000 |
| Max neural models | 100 | 500 |
| Compose model limit | 5 | 500 |
| Custom neural training | 10 hours per month | No limit, pay by the hour, 10 free hours each month |
| Classifier training pages | 25,000 | 25,000 |
| Classifier document classes | 1,000 | 1,000 |
A page is not always a page. For PDFs and TIFFs each page or image counts as one, with no character limit. For Word and HTML files Azure counts blocks of 3,000 characters, so a 7,000-character document bills as three pages. Each Excel worksheet and each PowerPoint slide counts as a full page on its own. Images embedded or linked inside Office and HTML files are not counted, because the Read and Layout models do not analyze them.
Barcodes, formulas, font properties, high resolution and language detection are separate add-on meters at $6.00 per 1,000 pages each. Switch on two and you have added $12.00 per 1,000 on top of whatever the base model costs.
Microsoft states that if your custom training dataset has no label files, the service generates them for you and bills you for Layout model usage. That is $10.00 per 1,000 training pages you may not have budgeted for, on top of training time.
A legacy meter called S0 Query Pages is live at $200 per 1,000 pages, while the current one, S0 pages for query fields, is $10. If query charges look absurd, check which meter caught them before concluding the feature is expensive.
The v3.1 documentation says training a custom model is always free. The v4.0 footnote says training beyond 20 runs in a calendar month bills on the training tier, and the quota table gives 10 free hours a month then $3.00 an hour. Both sentences are in Microsoft docs today.
One Excel worksheet is one page whether or not anyone opens it. A workbook with twelve monthly tabs is twelve pages, and hidden sheets are not excluded from the count.
REST API v2.1 reaches end of support on September 15, 2027, and REST API 2022-08-31 v3.0 on March 30, 2029. Current is v4.0, released as 2024-11-30 GA.
The arithmetic is shown so you can check it against the rate table above rather than take our word for it. All figures are Standard tier, East US, before any commitment.
| The job | Rate applied | Arithmetic | Monthly cost |
|---|---|---|---|
| 10,000 scanned pages, text only | Read at $1.50 per 1,000 | 10 x $1.50 | $15.00 |
| 10,000 pages, tables and structure | Layout at $10.00 per 1,000 | 10 x $10.00 | $100.00 |
| 10,000 invoices, one page each | Prebuilt invoice at $10.00 per 1,000 | 10 x $10.00 | $100.00 |
| 10,000 pages through your own trained model | Custom extraction at $30.00 per 1,000 | 10 x $30.00 | $300.00 |
| The same 10,000 pages on the legacy meter | Form Recognizer S0 Custom Pages at $50.00 | 10 x $50.00 | $500.00 |
| 250,000 pages of plain OCR, pay as you go | Read at $1.50 per 1,000, below the 1M break | 250 x $1.50 | $375.00 |
| 500,000 pages of plain OCR, Read 500K commitment | Commitment unit, $375 a month | Flat monthly unit | $375.00 |
| 1,000,000 pages, custom extraction, pay as you go | Custom at $30.00 per 1,000, no volume tier exists | 1,000 x $30.00 | $30,000.00 |
| The same 1,000,000 pages on the Custom 1000K commitment | Commitment unit, $18,000 a month | Flat monthly unit | $18,000.00 |
All three charge exactly $1.50 per 1,000 pages for plain OCR, so the entry rate is not a differentiator. The gaps open up on structured work, and on whether a capability exists at all.
| Per 1,000 pages | Azure | AWS Textract | Google Document AI | Read |
|---|---|---|---|---|
| Plain OCR / text detection | $1.50 | $1.50 | $1.50 | A three-way tie at the entry rate |
| Above 1,000,000 pages a month | $0.60 | $0.60 | $0.60 (above 5M) | Azure and AWS break at 1M, Google at 5M |
| Tables and layout | $10.00 | $15.00 | $10.00 | Azure and Google, 33% under AWS |
| Prebuilt invoice | $10.00 | $10.00 | $0.10 per 10-page block | Google bills invoices per block, not per page |
| Custom trained extraction | $30.00 | $50.00 (Forms) | $30.00 | Azure and Google tie, AWS Forms is the outlier |
| Document classification | $3.00 | Not offered | $5.00 | Azure is the cheapest, AWS has no classifier |
| Ask for a named field | $10.00 | $15.00 | Not offered as a meter | Azure query fields undercut AWS Queries |
| Batch discount | None | None | None | None of the big three discount batch work |
| Max pages in one call | 2,000 | 3,000 (async with splitter) | 15 online, 1,000 batch | Google has the lowest online ceiling |
Full workings for each vendor live on the cross-vendor OCR pricing per 1,000 pages reference, and the buying decision between them is laid out on the OCR API pricing comparison. If you are weighing Azure's newer service instead, the Azure Content Understanding vs Document Intelligence comparison covers where each one wins, and the per-call ceilings are on the Azure Document Intelligence limits page.
We are not going to pretend DocuOCR undercuts Azure on price per page. It does not. Azure Read is $1.50 per 1,000 pages and our plans work out to roughly $14 to $20 per 1,000. That comparison is real, and if your requirement is raw text off a scan at volume, the Azure API is the cheaper answer and you should use it.
The reason teams still buy a product instead is that the API rate is the smallest line in the total. Azure returns JSON. It does not queue the documents, decide which model each one needs, hold the ones that failed validation, give a non-engineer a screen to correct a misread field, keep an audit trail of who changed what, or push the finished rows into your accounting system. Building that around a $1.50 API is where the actual budget goes, usually in engineer months.
Azure Form Recognizer is now called Azure AI Document Intelligence, and the rates are unchanged by the rename. On the Standard (S0) tier you pay $1.50 per 1,000 pages for plain OCR with the Read model, $10.00 per 1,000 pages for Layout or any prebuilt model such as invoice or receipt, $30.00 per 1,000 pages for a custom model you trained yourself, and $3.00 per 1,000 pages for document classification. Read drops to $0.60 per 1,000 above one million pages a month.
Yes. Microsoft renamed Azure Form Recognizer to Azure AI Document Intelligence in July 2023, and the documentation now files it under Azure Document Intelligence in Foundry Tools. It is the same service, the same endpoints and the same models. The rename broke nothing, which is exactly why the old name is still everywhere: the SDK packages are still called Azure.AI.FormRecognizer, azure-ai-formrecognizer and @azure/ai-form-recognizer, and Microsoft still asks you to tag support questions azure-form-recognizer.
Because Microsoft never retired the old billing product. The Azure Retail Prices API still returns a live product called Form Recognizer with four meters of its own, alongside the current Azure Document Intelligence product with 68 meters. Both now sit under the service name Foundry Tools. Check the exact meter name on your invoice rather than the product heading, because the two products do not price every operation the same way.
They are $20 per 1,000 pages apart. Microsoft's price feed lists S0 Custom Pages at $50.00 per 1,000 under the legacy Form Recognizer product, effective August 2021, and at $30.00 per 1,000 under Azure Document Intelligence, effective June 2024. That makes the legacy meter about 67% more expensive for the same custom extraction work. We verified it in East US, West US 2 and North Europe. Read the meter name on your own invoice to see which one you are on.
No, and this is unusual. Every batch meter in Microsoft's price feed is priced identically to its synchronous twin: batch Read is $1.50 against Read at $1.50, batch Layout $10.00 against $10.00, batch custom extraction $30.00 against $30.00. The Batch API buys you convenience and throughput, not a lower rate. Mistral, Anthropic and OpenAI all cut batch work by roughly 50%, so if batch pricing matters to your budget, Azure is the wrong place to look for it.
Yes, and it turns cheaper far earlier than most people expect. Pay as you go for Read is $1.50 per 1,000 pages and only falls to $0.60 above one million pages a month. The Read 500K commitment tier is $375 a month for 500,000 pages, which works out to $0.75 per 1,000. Since 250,000 pages on pay as you go also costs $375, the commitment gives you twice the pages for the same spend at that point.
For PDFs and TIFFs, each page or image counts as one page with no character limit. For Word and HTML files, Azure counts blocks of 3,000 characters, so a 7,000-character document bills as three pages. Each Excel worksheet and each PowerPoint slide counts as a full page. Embedded or linked images inside Office and HTML files are not counted as extra pages, because the Read and Layout models do not analyze them.
There is a Free (F0) tier, but it is a testing tier rather than a usable free allowance. Microsoft states plainly that the free tier limits the analyze response to only the first two pages in a request, so any real document comes back truncated. F0 also caps you at 4 MB per file and one transaction per second, against 500 MB and 15 transactions per second on Standard. Use it to confirm the output shape, not to run a workload.
It depends on which API version you are on, and Microsoft's own documentation says both things in different places. The v3.1 billing section states that training a custom model is always free. The v4.0 footnote states that training requests over 20 in a calendar month are billed on the training tier, and the quota table gives Standard "no limit (pay by the hour), start with 10 free hours each month" at $3.00 per hour. On v4.0, budget for paid training once you go past 10 hours or 20 runs in a month.
It is a legacy meter called S0 Query Pages, and it is still live in Microsoft's price feed at $200 per 1,000 pages. The current query fields meter, S0 pages for query fields, is $10 per 1,000. That is a 20 times difference for what buyers think of as the same feature, so if you see query charges you did not expect, check which of the two meters your usage landed on before you assume the feature is simply expensive.
On pay as you go, yes: Microsoft's documentation states that container pricing is the same as cloud service pricing. The commitment tiers are a different story. The price feed lists Connected container commitment SKUs 15 to 20% below the equivalent Azure ones, so a Prebuilt 20K commitment is $152 a month on Connected against $190 on Azure, and a Custom 1000K commitment is $15,300 against $18,000. Both statements are true at the same time because they describe different purchase models.
Three catch people out. First, labeling: Microsoft states that if your custom training dataset has no label files, the service generates them and bills you for Layout model usage on top of training. Second, add-on capabilities are $6.00 per 1,000 pages each, and they stack, so turning on barcodes and high resolution together costs $12.00 per 1,000 on top of the base model. Third, Excel worksheets bill per sheet, including sheets nobody looks at.
Microsoft has published two dates. REST API v2.1 reaches end of support on September 15, 2027. REST API 2022-08-31 v3.0 reaches end of support on March 30, 2029. The current version is v4.0, released as 2024-11-30 GA. If you are still calling FormRecognizerClient or FormTrainingClient in your code, you are on a v2.x client and you are inside the first of those two windows.
For plain OCR all three charge exactly $1.50 per 1,000 pages, so there is nothing to choose between them at the entry rate. Azure pulls ahead on structured work: Layout is $10.00 against Textract Tables at $15.00, custom extraction is $30.00 against Textract Forms at $50.00, and Azure is the only one of the three with a document classifier, at $3.00 per 1,000. Google prices prebuilt invoices differently again, at $0.10 per 10-page block.
The current-name companion to this page, with the full model-by-model breakdown.
Microsoft's newer service, cheaper than Document Intelligence on every content extraction meter.
File size, page ceilings, transactions per second and what returns a 429.
The competing rate card, including why Forms at $50 is the outlier of the big three.
Processor by processor, including the per-block invoice rate.
Every vendor normalized to the same unit, so the numbers can be compared directly.
The buying decision rather than the rate card.
For teams who have decided the API is not the right shape for the problem.
Where per-page economics change, and what breaks first at scale.
A rate card cannot tell you whether the fields land correctly on your documents, and that is the only number that decides whether a pipeline is worth building. Upload one and read the output.
Azure rates on this page were read from Microsoft's Azure Retail Prices API on 10 August 2026, region East US. Rates change. Verify against the feed before you sign anything.
