Document AI Cost at Scale: Pages Are the Small Part

A document AI bill has two lines. The vendor line is pages times a per-page rate, and it is easy to find. The second line is people: the minutes someone spends checking what the machine was unsure about. At the volumes most teams start with, the second line is several times the first.

This post collects list prices from the official pages as of September 26, 2026, then runs a worked example through the document processing cost calculator to show where the money goes. It is part 9 of the Document AI series and builds on the review queue design, which is where most of that second line comes from.

What a page costs

The prices below are list prices for extracting fields from a document, not just reading its text, because that is what an invoice or form pipeline needs. All are US dollars, standard pay-as-you-go unless noted, read on 2026-09-26.

Sources: Datalab pricing and extraction modes, Extend pricing and credit docs, Google Document AI pricing, Azure Retail Prices API, and Mistral's API pricing.

A few details change the arithmetic:

• Datalab bills each processor separately. Its docs say extraction rates exclude parsing, which is billed when it runs, so a fast extraction is $6 plus $4 per 1,000 pages. The pricing page also says around 5% of balanced and accurate extractions bill above the flat rate, typically a dollar or two per document.
• Extend's extract includes a parse. Its pricing page says an extraction costs 3 credits a page plus 2 for the automatic parse, and gives the example of a 10-page Performance extraction at 50 credits, about $0.63. The Scale plan, $500 a month with 50,000 credits, drops additional credits to $0.01. Its Review Agent adds 1 credit a page.
• Azure's pricing page shows no numbers in its served HTML. On 2026-09-26 the Azure pricing page loaded its prices with JavaScript, so a plain fetch showed only the free tier of 500 pages a month. The figures above come from Microsoft's public Retail Prices API for East US. That is why the Lab calculator has no Azure preset. Enter the rate as a custom price.
• Google's specialized parsers price by document. The Invoice parser is $0.10 per count, where one count is up to 10 pages. For one-page invoices that is 10 cents a page. For ten-page ones it is one cent.

Commitment tiers change the picture at volume

Azure publishes monthly commitment tiers alongside pay-as-you-go. In the Retail Prices API on 2026-09-26, the custom extraction tier for 500,000 pages was $10,500 a month with overage at $21 per 1,000 pages. That is $0.021 a page against $0.030 pay-as-you-go. Google's custom extractor drops from $30 to $20 per 1,000 pages above a million pages a month. At a few thousand documents none of this applies. At a few hundred thousand, the tier you commit to matters more than which vendor's list price looked lowest.

Sending pages straight to an LLM

The other route skips the document AI vendor and sends the PDF to a general model with a JSON schema. The page cost is lower, but you pay in engineering: citations, confidence, verification and the review hooks above are yours to build.

Sources: Gemini API pricing and media resolution, Claude pricing and PDF support, read 2026-09-26.

The Gemini figure is 560 tokens a page at the default setting in its media resolution table for Gemini 3 models. Google's pricing page says the Flash input price doubles to $1.50 on January 1, 2027, and output is $3.75 per million tokens until then. Claude's PDF docs say each page typically uses 1,500 to 3,000 text tokens and is also converted to an image billed under the vision rules. For Haiku 4.5 an image is capped at 1,568 tokens, which gives the upper end of the range. Both exclude your prompt, your schema and the output. At an illustrative 400 output tokens per document, output adds about $1.50 per 1,000 documents on Gemini 3.8 Flash and $2 on Haiku 4.5.

These are the cheapest rows in the post by a wide margin. They are also the rows where the per-page price is least of the story, because the review load depends on how well the extraction is checked, and here nothing is checked unless you build it.

A worked example with the Lab calculator

The calculator adds two costs: vendor cost, which is documents times pages times price per page, and review cost, which is documents times the exception rate times minutes per exception times an hourly rate. It compares the total against doing every document by hand. Its defaults are 2,000 documents a month, 5 pages each, 8% exceptions, 6 minutes per exception, 4 minutes to do a document by hand, and $45 an hour. Those defaults are illustrative, not measured. The point is the shape of the result.

The figures below come from running the calculator's own calculation code with these inputs.

Doing all 2,000 documents by hand at 4 minutes each is $6,000 a month under the same inputs, so every scenario saves money. What changes the result is the exception rate, not the vendor price. Moving from the cheapest extraction to the most expensive in the table adds $525 a month. Moving the exception rate from 8% to 20% adds $1,080.

The break-even rule

Each percentage point of exception rate is worth documents times 1% times review minutes times the hourly rate. With the defaults, that is 2,000 times 0.01 times 0.1 hours times $45, or $90 a month per point.

That gives a direct test for any pricier extractor. The Extend Performance row costs $525 a month more than the Datalab row here. It pays for itself only if it cuts the exception rate by at least 525 divided by 90, about 5.8 points, from 8% to roughly 2.2%. You cannot know that from a pricing page. You find out by running both on the same documents and scoring them, as in measuring extraction accuracy field by field.

The rule scales. At 100,000 documents a month with the same inputs, the calculator gives $5,000 of vendor cost against $36,000 of review, and each exception point is worth $4,500. At that volume, a vendor that removes a single point of exceptions can justify a noticeably higher per-page rate.

Costs the calculator leaves out

• The audit sample. The review queue post recommends reviewing a random sample of auto-accepted documents. At a fixed 189 documents a month and the same 6 minutes and $45, that adds about $850. At small volume it can exceed the exception review.
• Building and running the pipeline. The calculator has a one-time build cost field for payback, but not ongoing maintenance, hosting or the n8n or cloud bill for the pipeline around the extractor.
• Reprocessing. A schema change that requires rerunning last month's documents bills those pages again.
• Errors that get through. A wrong total that reaches the ledger costs whatever it takes to find and reverse it. That is the cost the review queue exists to prevent, and it does not appear on any vendor's pricing page.

How to use this

1. Estimate pages per document and monthly volume from real traffic, not a guess.
2. Price the vendor line from the official page on the day you decide, and date it. Several of these prices carry announced changes.
3. Run a pilot on a few hundred of your documents to measure the exception rate and minutes per review for each candidate.
4. Put those numbers into the calculator and compare totals, not per-page prices.
5. Use the $-per-point rule to decide whether a pricier extractor earns its premium.

Previous in the series: designing the document review queue. Next: the Document AI glossary. For each vendor's rates in context, see Extend vs Datalab vs other parsers and Datalab in practice.