Azure Document Intelligence Pricing Calculator: Estimate Your Bill by Model and Volume

To estimate an Azure Document Intelligence bill you need exactly three numbers: which model you call, how many pages you send it in a month, and whether you run it in Azure's cloud or in a container. Multiply pages by the model's rate, then check whether a commitment tier beats pay as you go at your volume. Read costs $1.50 per 1,000 pages, Layout and every prebuilt model cost $10.00, and custom extraction costs $30.00. Everything else is a modifier on those three figures.

That sounds simple enough to do on the back of an envelope, and mostly it is. The reason estimates come in low is not the arithmetic. It is that the model you think you are calling is often not the one Azure bills you for, and that a handful of line items get switched on during a proof of concept and never switched off. This is the walkthrough, with the numbers pulled from Microsoft's public retail price feed in August 2026.

Step one: pick the right meter

This is where most estimates go wrong, so it is worth being careful. Azure does not charge for OCR as a single thing. It charges for the model you invoked, and the gap between the cheapest and the most expensive is twentyfold.

The single most common estimating mistake is budgeting Read and then calling Layout. People do it because they think of layout as a formatting option rather than a separate model, and because the code change is one parameter. On the invoice it is a different meter at 6.7 times the price. Microsoft bills Layout under the prebuilt line item, alongside invoice and receipt extraction, which is why asking for a table costs the same as asking for an invoice total.

Read is the only model with a volume break. Above one million pages a month it drops from $1.50 to $0.60 per 1,000. Nothing else tiers down at any volume.

Step two: multiply by your real page count, not your document count

Azure bills pages analyzed, and its own documentation is blunt about what that means: "Document Intelligence is billed by pages analyzed. When a document is submitted for analysis all pages are analyzed unless a page range is specified."

So a 40-page loan file where you only need the signature page costs 40 pages, not one, unless you pass a page range. Teams that estimate from document counts rather than page counts routinely land 5 to 10 times under. Count pages, and if your documents are long and you only need part of them, the pages parameter is the cheapest optimization available to you.

Three worked examples

An AP team running 8,000 invoices a month, averaging 2 pages. That is 16,000 pages through prebuilt-invoice at $10.00 per 1,000, so $160 a month. If they also classify a mixed inbox first, add 16,000 pages of classification at $3.00, which is $48. Total around $208 a month.

A records team digitizing 400,000 pages a month of plain text. All of it through Read at $1.50 per 1,000 comes to $600 a month. They are below the one million page break, so no discount applies. If someone switches this job to Layout because the output looked untidy, the same work becomes $4,000.

A team extracting 12 fields from a proprietary form, 60,000 pages a month. Custom extraction at $30.00 per 1,000 is $1,800 a month, plus training time at $3.00 per hour, which is negligible. This is the case where a commitment tier starts to matter.

Step three: check whether a commitment tier beats pay as you go

A commitment tier is a fixed monthly fee that buys a page allowance and bills whether you use it or not. The useful question is not what the tier costs, it is how much of it you have to consume before it saves anything.

Two things fall out of that table. The Read tiers are generous, because you only need to use 40 to 50 percent of the allowance to come out ahead. The small prebuilt tier is close to pointless: at $190 for 20,000 pages you need 19,000 of them every single month, and the best case saves 5 percent.

The risk is also one-sided, and in the direction people do not expect. We checked the published overage rate against the effective in-tier rate for every tier and both channels, 24 combinations in total, and all 24 match to the cent. Going over your commitment costs exactly what the commitment costs per page. There is no penalty rate. Every dollar of commitment risk sits on pages you paid for and did not send, which means the safe move is always to buy the tier below your forecast rather than the one above it.

The four line items that break a clean estimate

These are the ones that show up on the invoice and not in the spreadsheet.

• Add-on features, $6.00 per 1,000 pages. High Resolution, Font and Formula are request flags. They are easy to enable while testing a difficult scan and easy to leave on. They bill on top of the model you called, so a prebuilt page with an add-on is $16.00 per 1,000, not $10.00.
• Query fields, $10.00 per 1,000 pages. Also charged on top. A prebuilt invoice call with query fields doubles to $20.00 per 1,000.
• Whole-document billing. Covered above, and the biggest single distorter of an estimate built from document counts.
• Reprocessing. Every retry is a fresh billable page. If 3 percent of your documents fail validation and get resubmitted, your real page count is 3 percent higher than your input count, and if a bad batch gets rerun wholesale it is much higher than that.

Notably absent from that list is batch. Azure prices batch identically to synchronous on every paired meter, so building a queue to chase a batch discount is engineering effort spent on a saving that does not exist. Region is absent too: every commercial Azure region charges the same rate, and the only exception anywhere is Azure Government at exactly 1.25 times commercial.

If you are considering containers

Running the model in a container you host changes the price, and by a different amount for each model family. Compared with the equivalent Azure cloud tier, a connected container is 10 percent cheaper for Read, 15 percent cheaper for custom extraction and 20 percent cheaper for prebuilt models. That holds exactly across all twelve published tiers, and it means containers make the weakest financial case for cheap high-volume text and the strongest for prebuilt extraction at volume.

Disconnected containers, the genuinely air-gapped option, are priced per year at exactly twelve times the connected monthly fee. There is no prepayment discount. And there is a procurement detail worth knowing before you promise a number to a security team: the smallest 20,000-page disconnected tiers are listed in only three regions worldwide, none of them in the United States. A US team that needs an air gap starts at the 100,000-page tier, so the floor is $8,640 a year for prebuilt models and $24,480 a year for custom extraction. The full rate tables, including every meter name and effective date, are on our Azure Document Intelligence pricing breakdown.

What the calculator will never tell you

Every number above is an API cost. A working document pipeline also needs the classification step that routes a mixed inbox, the validation rules that catch a misread total before it reaches your ledger, a review queue for anything under your confidence threshold, storage, and the engineers who keep all of it running. For most teams that surrounding work costs considerably more than the extraction itself, which is the honest comparison we lay out in our OCR API pricing comparison.

It is also worth separating the estimate from the actual. An estimate is a forecast made once; a document pipeline's page volume drifts with your business, and the meter you are billed on can change with a one-line code edit nobody flags in review. Connecting your subscription to something that shows what each cloud service actually cost you last month turns that drift into something you notice in week one rather than at renewal. If you want the same treatment applied to Amazon, our AWS Textract pricing analysis covers the meters and the region trap on that side.

The short version

Pages times the rate of the model you are actually calling, plus $6.00 per 1,000 for any add-on and $10.00 per 1,000 for query fields, plus a few percent for retries. Compare that against the commitment tier at your volume and buy the tier below your forecast rather than above it. Ignore batch and ignore region, because neither changes the price. Then, before you commit to any of it, run the documents you actually have through an extraction and see how many fields a person still has to fix, because that number decides your real cost far more than $1.50 versus $10.00 ever will.