Microsoft's newest document AI service is cheaper than Document Intelligence on every content extraction meter: $1.00 per 1,000 pages for OCR against $1.50, and $5.00 for layout against $10.00. The rates below were read off Microsoft's official retail price feed, not a marketing page.
Upload a document to extract
Drop files here or click to upload
Up to 50 files
Uploading...
A per-page rate cannot tell you whether the fields come out right on your documents. Drop one in and look at the output before you commit to a vendor.
Azure Content Understanding charges $1.00 per 1,000 pages to OCR a scanned document, $5.00 per 1,000 pages to run layout analysis with tables and structure, and $0.01 per 1,000 pages for digital files such as Word, Excel and email that need no OCR at all. Those three rates are the Basic, Standard and Minimal meters. Every one of them undercuts Azure AI Document Intelligence, which charges $1.50 for the same OCR, $10.00 for the same layout, and has no digital-file meter at all. If you also switch on generative field extraction, add $1.00 per 1,000 pages of contextualization plus the input and output tokens of the Foundry model you connect, billed separately to that deployment. Microsoft's own worked example for invoices on a GPT-4.1-mini deployment lands at $8.37 per 1,000 pages.
Field extraction has no flat rate. It rides on a model deployment you own and bill separately.
Microsoft's marketing pricing page is not the authoritative source. The Azure Retail Prices API is: it is the feed the portal bills from. These are the seven meters it returns for Content Understanding in East US, with the date each rate took effect, plus the one input type Microsoft charges nothing to extract.
| Meter | What it covers | Unit | Official rate | Effective |
|---|---|---|---|---|
| Doc Content Extraction Minimal | Digital files: DOCX, XLSX, PPTX, HTML, TXT, MSG, EML | Per 1,000 pages | $0.01 | 2026-01-01 |
| Doc Content Extraction Basic | OCR on image-based files: scanned PDF, TIFF, JPG, PNG | Per 1,000 pages | $1.00 | 2025-12-01 |
| Doc Content Extraction Standard | Layout analysis: tables and structure, image-based files | Per 1,000 pages | $5.00 | 2025-12-01 |
| Add-On Layout Pages | Layout add-on | Per 1,000 pages | $5.00 | 2025-12-01 |
| Std Contextualization Tokens | Schema formatting, confidence scores, source grounding | Per 1M tokens | $1.00 | 2025-12-01 |
| Audio Content Extraction | Speech to text | Per hour | $0.36 | 2025-12-01 |
| Video Content Extraction | Frame extraction, shot detection, transcription | Per hour | $1.00 | 2025-12-01 |
| Image content extraction | Not charged. Images incur no content extraction fee at all. | Per image | $0.00 | No meter exists |
Read from the Azure Retail Prices API on July 13, 2026, product "Azure Content Understanding", region East US. Rates vary by region and Microsoft changes them without notice.
This is the part of Content Understanding pricing that catches people out, and it works in your favor. You are charged for the processing the service actually performed, not for the analyzer you called. Point a layout analyzer at a Word document and Azure still only bills the Minimal rate of $0.01 per 1,000 pages, because no OCR and no layout detection was needed to read a digital file.
The practical consequence is that your bill is driven by what your documents are, not by how you configured the call. A pipeline that is 80% digital PDFs and Word files costs almost nothing to extract, no matter which analyzer you point at it. A pipeline of scanned paper pays the Basic or Standard rate on every page.
Always the Minimal meter, whatever analyzer you call. No OCR is performed, so none is billed.
The Basic meter. OCR runs, but no table or structure detection.
The Standard meter. Table recognition and structural elements on a scanned file.
All rates per 1,000 pages. Images, as opposed to image-based documents, carry no content extraction charge at all.
On content extraction, yes, on every single meter. On structured field extraction it depends entirely on which model you attach, and on two counts Document Intelligence still wins outright. Both sets of rates were pulled from the same Microsoft price feed on the same day.
| The job | Content Understanding | Document Intelligence | Winner |
|---|---|---|---|
| Plain OCR on a scanned PDF | $1.00 (Basic) | $1.50 (Read) | Content Understanding, 33% cheaper |
| Layout, tables and structure | $5.00 (Standard) | $10.00 (Layout) | Content Understanding, 50% cheaper |
| Digital DOCX, XLSX, HTML, TXT | $0.01 (Minimal) | $1.50 (Read, no digital meter) | Content Understanding, 150x cheaper |
| A standard invoice into fields | About $8.37 (GPT-4.1-mini) | $10.00 (prebuilt invoice) | Close. Content Understanding on a mini model |
| Custom layout into fields | $6.00 plus your model tokens | $30.00 flat (custom extraction) | Content Understanding on a mini model, but variable |
| Document classification | Contextualization plus model tokens | $3.00 flat (classifier) | Document Intelligence, cheaper and simpler |
| Max pages in one call | 300 pages | 2,000 pages | Document Intelligence, 6.7x higher ceiling |
| Predictable flat rate | No, generative cost varies by model | Yes, published per-page rates | Document Intelligence |
| Batch discount | None | None | Neither |
Rates per 1,000 pages unless stated. Full detail on the older service is on our Azure Document Intelligence pricing breakdown, and the two services are compared feature by feature on Content Understanding vs Document Intelligence.
The headline meter is only the first line of the bill. These are the fully worked totals Microsoft publishes in its own pricing documentation, including contextualization and the Foundry model tokens.
| Workload | Setup | Extraction + model + contextualization | Total |
|---|---|---|---|
| 1,000 invoice pages into structured fields | GPT-4.1-mini, source grounding and confidence on | $5.00 + $2.08 + $0.29 + $1.00 | $8.37 |
| The same 1,000 invoice pages | Full GPT-4.1 instead of mini | Microsoft's stated note | About $33.00 |
| 1,000 pages for a RAG index, with figure analysis | GPT-4.1 global deployment | $5.00 + $4.00 + $3.20 + $1.00 | $13.20 |
| 1,000 images captioned | GPT-4.1, no content extraction charge on images | $0.00 + $2.09 + $1.36 + $1.00 | $4.45 |
| One hour of call center audio | GPT-4.1-mini | $0.36 + $0.01 + $0.10 | $0.47 |
Look at the first two rows. The same 1,000 invoice pages cost $8.37 on a GPT-4.1-mini deployment and roughly $33.00 on a full GPT-4.1 deployment. That is a four-fold swing driven entirely by a model choice, and it is the difference between beating Document Intelligence custom extraction ($30.00 flat per 1,000 pages) and losing to it. Microsoft's own guidance is blunt about this: a mini model can cut generative cost by up to 80%, and content extraction and contextualization charges do not change at all when you switch models.
One thing Microsoft's cost examples do not flag: every one of them is priced on the GPT-4.1 family, and that family is scheduled for retirement in October 2026. The supported replacement in the Content Understanding docs is gpt-5.2. So if you are budgeting a Content Understanding pipeline today, the extraction and contextualization meters on this page are stable, but the model half of the bill is going to be re-priced on a model that does not appear in a single one of Microsoft's published examples. Price the meters from this table and price the tokens from whatever model you actually deploy.
Generative features run on a Foundry model deployment that you own. Those input, output and embedding tokens are billed to that deployment, not to Content Understanding. The per-page rate on this page is real, but it is not the whole invoice.
Microsoft counts one worksheet as one page, and its documentation states that includes hidden sheets. One slide is one page. For text, HTML, XML and email, every 3,000 characters counts as a page, rounded up.
Content Understanding accepts 200 MB and 300 pages in a single document. Document Intelligence accepts 500 MB and 2,000 pages. If you push long files, that gap matters more than the rate.
Microsoft does not bill content extraction or contextualization when a request fails with an error such as a 400. If a model call already succeeded before the failure, Foundry still bills those tokens.
Source grounding and confidence scores roughly double token usage. So does adding training examples, and so does segmentation or categorization. Extractive mode adds about 50%. The extraction meter stays flat while the model bill climbs.
API versions 2024-12-01-preview and 2025-05-01-preview are being retired. The generally available version is 2025-11-01. Anything still pointing at a preview endpoint needs to move now.
Every vendor normalized onto one unit, including the per-document and per-token ones.
Read itThe feature-by-feature decision, the migration path and where the older service still wins.
Read itFull rate card, commitment tiers Microsoft does not publish, and the batch finding.
Read itThe buyer pillar: which OCR API to actually pick, by workflow.
Read itOfficial rate card, the 3-month free tier, and the models Textract does not have.
Read itEvery meter including the ones most guides miss, and the idle hosting fee.
Read itContent Understanding at $1.00 per 1,000 pages is genuinely the cheapest way to get text out of a scan on Azure. What it hands back is an API response. You still have to build the classification that routes each document to the right analyzer, the review screen where a human fixes the fields the model got wrong, the validation rules, and the export into your accounting system or ERP. On most projects that engineering, not the meter, is the real cost.
DocuOCR prices higher per page, roughly $14 to $20 per 1,000, because all of that is already built. Whether that trade is worth it depends on how much of the pipeline you want to own.
No rate table tells you whether the fields come out right on the forms you actually process. Upload one and compare the output yourself.
Extract a document freeLast updated July 2026. Rates read from Microsoft's official Azure Retail Prices API and re-verified before publication.
Z tej samej rodziny narzędzi
