What Is Azure Content Understanding?

Azure AI Content Understanding is Microsoft's newer content extraction service. It takes unstructured input, a scanned PDF, a Word file, a photo, a call recording, a video, and returns structured output that matches a schema you define in JSON. It reached general availability with API version 2025-11-01, and unlike Azure AI Document Intelligence, which only reads documents, it covers four modalities at once.

The short version for anyone comparing the two: Content Understanding is cheaper than Document Intelligence on every content extraction meter, it handles audio and video that Document Intelligence cannot touch at any price, and it runs generative work on a model deployment you provide rather than on a model Microsoft trained. That last point is the one that changes how you budget, and it is the part most write-ups skip.

What does Azure Content Understanding actually do?

It performs two separable jobs, and you can pay for the first without the second.

Content extraction turns the raw input into text and structure. For documents that means OCR and layout analysis. For audio it means speech to text with speaker diarization. For video it means frame sampling at roughly one frame per second, shot detection and a transcript. This is billed per 1,000 pages for documents and per hour for audio and video, and it is the only charge if you stop here.

Generative features sit on top: field extraction into your schema, figure analysis, segmentation, categorization and training on labeled examples. These call a large language model, and here is the structural difference from every other Azure document service. You bring your own Microsoft Foundry model deployment. Content Understanding orchestrates the call, but the input and output tokens are billed to your deployment, not to Content Understanding. Microsoft currently supports GPT-5.2 and the GPT-4.1 family for completion, plus the text-embedding-3 models for embeddings.

How much does Azure Content Understanding cost?

Document content extraction has three meters, and Azure picks which one fires based on the work it actually performed, not the analyzer you called. Rates below were read from Microsoft's official Azure Retail Prices API in July 2026 for the East US region.

Audio content extraction is $0.36 an hour and video is $1.00 an hour. Images carry no content extraction charge at all. Microsoft's pricing documentation labels the audio rate per minute in one example but then does the arithmetic as 60 minutes for $0.36, and the price feed settles it: the meter's unit is one hour.

The rule about which meter fires is worth reading twice, because it works in your favor. Microsoft's documentation says you are charged for the type of processing performed, not the analyzer you select. Point a layout analyzer at a Word document and you still pay only the Minimal rate of $0.01 per 1,000 pages, because no OCR and no layout detection was required to read a digital file. Your bill is driven by what your documents are, not by how you configured the call.

Is Azure Content Understanding cheaper than Document Intelligence?

On content extraction, yes, on every meter. Plain OCR is $1.00 per 1,000 pages against $1.50 for Document Intelligence Read, so 33% cheaper. Layout is $5.00 against $10.00, so 50% cheaper. Digital files are $0.01 against $1.50, because Document Intelligence has no digital meter and charges the full Read rate to parse a file that was never a scan. That is a 150-fold difference, and for a pipeline heavy in Word, Excel and email it is by far the largest saving available.

Structured field extraction is the exception, and it is the exception that decides most migrations. Document Intelligence charges a flat $30 per 1,000 pages for custom extraction and bills you nothing else. Content Understanding has no flat rate for the same job: you pay $5.00 content extraction, $1.00 contextualization, and then whatever your model costs. Microsoft's own worked example for invoices on a GPT-4.1-mini deployment totals $8.37 per 1,000 pages. Their own note adds that swapping mini for a full GPT-4.1 deployment lifts the same job to roughly $33.

Read that as the real finding: the model you attach, not the service you choose, decides whether you beat Document Intelligence by 3.6x or lose to it slightly. We break the full arithmetic down on our Azure Content Understanding pricing reference.

Is Content Understanding replacing Document Intelligence?

No. Microsoft has announced no end of support for Azure AI Document Intelligence, and every one of its billing meters is live in the official price feed, including some that took effect as recently as January 2026. Both services are sold and supported. Posts declaring Document Intelligence dead are guessing.

There is one real, dated deadline, and it belongs to Content Understanding rather than to the older service: API versions 2024-12-01-preview and 2025-05-01-preview are being retired on July 15, 2026. The generally available version is 2025-11-01. If your code still targets a preview endpoint, it needs to move, and because the GA release also changed how generative features are billed, your old cost estimate probably no longer holds.

What are the Azure Content Understanding limits?

This is where the newer service is genuinely weaker, and it catches people who process long files.

• 300 pages and 200 MB per document. Document Intelligence accepts 2,000 pages and 500 MB, the highest single-call ceiling of any major vendor.
• 1 million characters for digital and text files, and 1 MB for plain text or email.
• 1,000 pages or images a minute and 3,000 operations a minute on the standard S0 tier.
• Pro mode is tighter still: 100 MB, 150 pages, and it only accepts PDF, TIFF and image files.

There is also a billing detail hiding in the page count. For files that have no natural pages, Microsoft uses page-equivalent rules: 3,000 characters counts as one page for text, HTML, Markdown, XML and email; one Excel worksheet counts as one page, including hidden sheets; one PowerPoint slide counts as one page; and Word documents use their native pagination. The hidden-worksheet rule is the one that produces a surprising invoice.

What Content Understanding does that Document Intelligence cannot

Four things stand out beyond the price.

Audio and video. Document Intelligence has no capability here at all. If call recordings sit in the same pipeline as your paperwork, one service now covers both.

Confidence scores and source grounding as first-class output. Every extracted field can be traced back to where it appeared on the page, and each carries a confidence value. If a human reviews low-confidence fields, that signal is what decides who sees what.

In-context learning. You can improve extraction quality on a new template by supplying a handful of labeled examples, without retraining and redeploying a model. For documents that vary by supplier or by state, that iteration loop is dramatically shorter than the train-and-deploy cycle a custom Document Intelligence model requires.

A schema instead of a model. You describe the fields you want in JSON, up to 1,000 of them, including nested groups and tables. That is a different mental model from picking a prebuilt model and accepting its fixed schema.

Where Document Intelligence still wins

Three places, and they are not small. It takes 2,000 pages in a single call against 300. It charges a flat, forecastable $30 per 1,000 pages for custom extraction with no second bill to reconcile. And its document classifier is $3 per 1,000 pages, flat, where categorization on Content Understanding becomes contextualization plus model tokens on every page you route. For high-volume classification of long documents on a fixed budget, the older service is still the cleaner answer. We lay the whole decision out on Content Understanding vs Document Intelligence.

Does Azure charge for failed requests?

Not for content extraction or contextualization. Microsoft states that a request failing with an error such as a 400 is not billed for either. The caveat is that if a Foundry model call already succeeded before the failure, those tokens are still billed by the model deployment under Foundry's own rules. Google Document AI has the same policy on failed requests; AWS Textract does not publish one as clearly.

What it does not give you

Content Understanding returns a JSON response. It does not route each document to the right analyzer, it does not give you a screen where a person corrects the fields the model got wrong, it does not enforce your validation rules, and it does not push the result into your accounting system. Those are the parts you build, and on most projects that engineering costs more than the meter ever will. If the documents in question are employee receipts, for instance, a platform that already reads and categorizes them end to end closes a loop that a raw extraction API leaves open.

That is the honest trade. An Azure API at $1.00 per 1,000 pages is the cheapest way to get text off a scan, and it is a component, not a pipeline. A finished product prices higher per page precisely because the workflow is already there.

Should you use it?

Use Content Understanding if you are paying for OCR or layout at volume, if a meaningful share of your input is digital rather than scanned, if your layouts are irregular enough that trained models keep missing, or if audio and video belong in the same pipeline. Stay on Document Intelligence if you need a flat forecastable rate, if you send documents longer than 300 pages, if you route high volumes through the classifier, or if the prebuilt models already return what you need.

And if you would rather not own the pipeline at all, upload a document to the tool at the top of this page and compare what a finished extraction actually returns.

Rates and limits verified July 2026 from Microsoft's Azure Retail Prices API and Microsoft Learn. Microsoft changes rates without notice; re-check before you budget against them.