Document AI Pricing per Page: Azure vs AWS Textract vs Google, at 1,000 to 100,000 Pages

Intelligent document processing sounds like one product. It's really a price list with a dozen lines. The line you pick decides the bill. Reading the text off a page and pulling the total off an invoice differ by up to 66 times on the same cloud.

So we read the price data from each vendor on 3 October 2026: the AWS Price List API, Azure's Retail Prices API and Google's pricing page. We then priced three document types at 1,000, 10,000 and 100,000 pages a month, and added one LLM-vision route using published token prices. We haven't run any of these services ourselves, and none of the vendors pay us. For the same kind of math on chat and phone AI, see our AI customer service cost guide and our voice agent per-minute breakdown.

What 1,000 pages costs, feature by feature

The cheapest line on every cloud is plain OCR at $1.50 per 1,000 pages. Everything that understands structure costs 2.7 to 66 times more.

Each number in that table traces to a vendor source, and the details below show how we read each one.

• AWS Textract. The AWS Price List API for us-east-1, published 11 September 2026. Forms plus Tables is the sum of the two lines, which is how Textract's own page writes it. AnalyzeDocument output "comes with OCR included irrespective of feature type selected", so you don't pay for OCR twice. Combined features such as Forms plus Queries have their own prices.
• Azure. The Retail Prices API. Azure's page groups "Document, Layout, Receipt, Invoice, ID, W-2, 1098 Tax forms, Health insurance card, Contract" under one "All Prebuilt Models" price, $10 per 1,000 pages. Read is $1.50. Custom extraction and custom generative extraction are $30, custom classification is $3, and the High Resolution, Font and Formula add-ons are $6 on top. Azure describes Layout as extracting "text and structure (tables, titles, paragraphs, selection marks)". It doesn't say key-value pairs, so we price form fields on the custom model.
• Google. The pricing page. Enterprise Document OCR, Form Parser, Layout Parser and the custom extractor are per page. The specialized parsers for invoices, expenses, utility bills, US driver licenses and US passports are per "count", and "1 count equals up to 10 pages in a document".

One caution on reading the table across rows. Azure Layout at $10 and AWS Forms plus Tables at $65 aren't the same product. Layout returns structure, while AWS Forms returns labeled key-value pairs. We haven't tested either one's output, so the price ratio is not a quality ratio.

Worked cost at 1,000, 10,000 and 100,000 pages

At 100,000 pages a month, plain OCR costs about $150, forms with tables $1,000 to $6,500 and invoices $1,000 to $10,000. The formula is monthly cost = pages ÷ 1,000 × price per 1,000 pages.

Our assumptions:

• One page per document. Real invoices run longer. On Azure and AWS the bill scales with pages. On Google a 3-page invoice is still $0.10, so the parser matches Azure and AWS ($0.01 a page) only at 10 pages per document.
• Free allowance. We subtracted only Google's, because it repeats monthly: the first 1,000 Enterprise OCR pages are free. Azure's free tier and AWS's new-customer allowance are in the next section.
• Volume tiers. None applies. AWS and Azure discount only above 1 million pages a month (Azure Read drops to $0.60, AWS DetectDocumentText to $0.60, AWS Forms to $40). Google's discount for OCR starts at 5 million pages. The one early break is AWS AnalyzeID, which falls from $25 to $10 per 1,000 above 100,000 pages.

For invoices, 100,000 pages cost $1,000 on Azure or AWS and $10,000 on Google. If your invoices are single pages, that's a $9,000 gap in the monthly bill from choosing a cloud on brand alone.

The LLM-vision route: cheap on paper

Sending each page to a vision LLM costs about $1 to $3 per 1,000 pages on a low-cost model and $7 to $14 on a mid-tier one. We priced it from published token rates, not from a test, so it carries no accuracy claim.

The formula is cost per 1,000 pages = 1,000 × (input tokens × input price + output tokens × output price) ÷ 1,000,000.

The assumptions are ours, so check them against your documents:

• Input tokens. Google's docs say each PDF page "is equivalent to 258 tokens", and we add 300 tokens for the instruction and the output schema, which makes 558. Anthropic's vision docs put a 1.19-megapixel page image at 1,521 tokens; with the same 300 added, that's 1,821.
• Output tokens. 700 for a page of plain text, 1,000 for JSON with tables, 300 for invoice fields. These are estimates, not measurements. Thinking tokens are billed as output on Gemini, so a model run with heavy reasoning costs more.
• Digital PDFs. Gemini's docs say you aren't charged for tokens from native text in a PDF. A PDF with a real text layer can cost less than our figures; a scan can't.

What the price leaves out is the part that matters. A managed extraction service returns what it read, usually with confidence scores. A language model can return a plausible number that isn't on the page. We haven't measured error rates on any route here, so a fair read is that the LLM price is a floor. A validation pass, a second model or a human check on totals and dates adds cost. Our explainer on what a token is shows why a longer prompt and a longer answer each add to every page.

Open-source and self-hosted OCR

Self-hosted OCR has no per-page fee, but we can't give you a per-page price for it. We found no sourced throughput figure for a self-hosted vision model, so we haven't priced the hardware.

Three tools are open and free to run, with licences checked on GitHub on 3 October 2026:

• Tesseract, Apache-2.0, a long-standing text OCR engine.
• PaddleOCR, Apache-2.0.
• Docling, MIT, which converts documents including PDFs into structured output.

The cost moves from a price list to your own time and servers: setup, a machine, updates, and checking the output. If you want GPU prices for running a vision model yourself, our self-hosted LLM break-even guide has rental rates, and our private LLM cost guide for EU companies covers the keep-it-in-your-own-region version. Self-hosting is mainly a data-control decision, not a savings one, at the volumes we priced.

Free tiers, minimums and commitments

None of the three clouds charges a monthly minimum on pay as you go, and each has a small free allowance. The allowances differ in how long they last.

• Azure. The free instance (F0) allows 500 pages a month, with no premium features and no Query Fields. It doesn't expire.
• AWS. The free tier lasts three months for new AWS customers. It covers 1,000 pages a month of DetectDocumentText, 100 pages each of Forms, Tables and Layout, 100 of Queries and their combinations, 100 each of AnalyzeExpense and AnalyzeID, and none for Custom Queries.
• Google. The first 1,000 Enterprise Document OCR pages each month are free, with no time limit. Other processors have no free line.
• Billing quirks. Azure bills "by pages analyzed", and all pages in a file are analyzed unless you pass a page range. Google doesn't bill failed requests (4xx or 5xx). AWS says custom pricing for high volume is by sales quote.

Commitments exist for steady volume. Azure's monthly tiers are the ones we could read in full: prebuilt models 20,000 pages for $190 a month (against $200 pay as you go), 100,000 for $900, 500,000 for $4,000, and 1 million for $7,500. Read is 500,000 pages for $375. Overage is billed per 1,000 pages at a lower rate. That's a 5% to 25% discount on prebuilt models for committing to the volume. Google has hosting fees on custom processors, $0.05 an hour for a deployed version, which is $438 a year for one.

Regions and EU prices

Frankfurt and West Europe cost the same as US East on the lines we checked. AWS's Frankfurt Textract prices match us-east-1 for text, tables, forms, expenses and IDs. Azure's West Europe meters match East US for Read, prebuilt, custom and add-ons. US government regions on Azure cost about 25% more.

Google's page shows one price list and no regional breakdown, so we haven't verified whether any region costs more and we don't print one. If you're keeping documents in the EU for compliance, the price isn't the question; where processing happens and who is a sub-processor are. Our EU data residency cost guide walks through that for AI APIs.

Which feature do you need?

Pick the cheapest feature that returns the field you need. Most overspend comes from using a forms feature to read plain text, or an invoice parser on a document that isn't an invoice.

Three rules of thumb follow from the prices:

• Searchable archive or full-text search. Plain OCR at $1.50. Anything more is paying for structure you won't use.
• Standard invoices, receipts and IDs. Azure's prebuilt models or AWS's AnalyzeExpense at $10 a page. On Google, check the page count first.
• Your own forms. A custom model costs $25 to $30 per 1,000 pages on all three clouds, plus training. Azure's custom template training is free, and neural training is free for the first 10 hours, then $3 an hour. AWS Custom Queries adapters train free.

Costs the per-page price doesn't show

The per-page price covers the extraction call. The rest of the bill is usually storage, orchestration and review.

• Retries and long files. Azure bills every page of a file unless you pass a page range, so a 200-page PDF where you need page 3 costs 200 pages.
• Add-ons. Azure's High Resolution, Font and Formula add-ons are $6 per 1,000 pages on top. Google's OCR add-ons are $6 too.
• Splitting and classifying. Mixed batches need a classifier first. Azure custom classification is $3 per 1,000 pages. Google's custom classifier and splitter are $5 per 1,000 up to 1 million pages.
• Custom model hosting. Google charges $0.05 an hour for each deployed custom processor version.
• Human review. A person checking the totals on a sample is the real cost of any route, and no price list shows it.

For the wider set of automation costs we've priced, start at the AI automation hub. aliteq sells none of these products and earns nothing from these links. Every price was read from the vendor's own price data or page on 3 October 2026.