OCR API Pricing: Document AI, AWS Textract and Azure Costs Per Page

The AWS, Azure and Google rates below were re-verified against the AWS Price List API, the Azure retail prices API and Google Cloud pricing.

OCR API pricing comparison table

List prices in US dollars, standard pay-as-you-go tier, US regions (Textract figures are US West Oregon, Azure figures are East US S0). Committed-use and enterprise contracts are lower and are not shown, because vendors do not publish them.

Two newer document APIs built on large models are priced on a different basis again. Mistral OCR API pricing is $4 per 1,000 pages, or $2 through its Batch API, for layout-aware text and tables without receipt fields. Reducto pricing is $10 per 1,000 pages to parse and $20 to extract fields to a schema you write. Both are verified from the vendors own pages. LlamaParse pricing runs on credits at $1.25 per 1,000, which is $10 per 1,000 pages for Cost-effective field extraction and $31.25 for Agentic, and Landing AI pricing is $0.01 per credit with a per page plus per character formula. Extend AI pricing is $0.0125 per credit, which works out to $15 per 1,000 pages for Light extraction and $62.50 for Performance extraction. Datalab pricing is the lowest published extraction rate in this group at $6 per 1,000 pages on its turbo tier, with a recurring $20 monthly free allowance.

How much does an OCR API cost per page?

A plain text OCR API costs about $0.0015 per page, or $1.50 per 1,000 pages, on AWS Textract and Azure. Structured receipt and invoice extraction costs about $0.01 per page on Textract and Azure, and $0.10 per document on Google Document AI. Custom trained models cost around $0.03 per page. In other words, asking for fields instead of text multiplies the bill by roughly seven.

The per-document trap in Google pricing

Google is the outlier, and the difference is easy to miss. Its expense and invoice parsers charge $0.10 per document, and one billed count covers up to 10 pages. If your documents are ten-page invoices, that is $0.01 per page and Google is competitive. If your documents are single-page coffee receipts, that is $0.10 per page, ten times what Textract and Azure charge for the same job. Average pages per document is therefore the single biggest driver of which API wins on price for you, and it is the number most comparison articles never mention. The full Google rate card, every processor meter and the exact count rounding examples Google publishes, is on our Google Document AI pricing reference.

What OCR costs at scale: 10,000 receipts a month

Per page rates are hard to feel. Here is the same workload priced on every API: 10,000 single page receipts a month, structured field extraction rather than plain text, at published list prices. Single page receipts are deliberately the worst case for per document billing, which is exactly the point, because most receipt workloads look like this.

Three things fall out of that table. Textract and Azure are the cheapest raw list prices at this volume and are also the two that hand you a JSON blob you still have to normalize, host, monitor, and map to your chart of accounts. Google is ten times more expensive than Textract on single page receipts purely because of the per document billing unit, and would be competitive on ten page invoices. Nanonets looks cheap per block and is the most expensive option here once you count the four to six blocks a real document workflow consumes.

The number the table cannot show is engineering. A cloud OCR API is a component, not a product: the cost of building the validation, retry, review, and export layers around it is usually larger than the API bill in year one, and it recurs every time a vendor changes an output format. That is the honest reason a bundled tool can cost more per page and still be cheaper overall for a team that is not staffing a document pipeline. If you are weighing a language model instead of a per page API, the billing units differ again and are compared on our LLM OCR page.

Where the cloud APIs are genuinely cheaper than we are

We should say this plainly, because a comparison that only flatters the author is not worth reading. If you need raw text off a page and nothing else, Textract DetectDocumentText at $1.50 per 1,000 pages is roughly thirteen times cheaper than our effective per-page cost, and you should use it. If you are extracting millions of pages a month and already have an engineering team maintaining the pipeline, per-page cloud pricing with volume discounts will beat any flat plan. Our receipt OCR API makes sense in the middle: you want vendor, date, tax, line items, and totals as clean fields, at a volume in the thousands rather than the millions, without paying an engineer for a month to build the normalization layer.

What is the cheapest OCR API?

For plain text, Google Enterprise Document OCR is cheapest because the first 1,000 pages each month are free, then $1.50 per 1,000. For structured receipts and invoices, Amazon Textract AnalyzeExpense and Azure prebuilt models tie at $0.01 per page. For very low volume, Veryfi is effectively free at 100 documents a month. The cheapest option changes with volume, so run your own page count through each row of the table above.

Volume tiers and where they kick in

Volume discounts on the big three start higher than most teams reach. Textract drops from $0.0015 to $0.0006 per page for raw OCR and from $0.01 to $0.008 for AnalyzeExpense only after one million pages in a month. Google drops Enterprise Document OCR from $1.50 to $0.60 per 1,000 only above five million counts, and drops the custom extractor from $30 to $20 per 1,000 above one million. Azure sells committed tiers instead, where you prepay a monthly block and pay a lower overage rate. Below about 100,000 pages a month, assume you are paying list price everywhere.

Costs that never appear on a pricing page

Three line items regularly double a projected OCR budget. The first is idle hosting: a Google custom processor bills $0.05 per hour while deployed, whether or not it processes anything, which is about $438 over a year. The second is training: Azure gives you 10 free hours of custom neural training on v4.0 and then charges $3 per hour. The third, and by far the largest, is engineering. A raw API response is a bag of key-value pairs with confidence scores, not a spreadsheet row. Someone has to map fields, split multi-page files, decide what happens when confidence drops below threshold, and build the export. Two developer weeks at a loaded US rate costs more than 500,000 pages of Textract AnalyzeExpense.

Do OCR APIs charge for failed requests?

Generally no. AWS does not bill for requests that return 4xx or 5xx errors, and Google states the same for failed requests. What you do pay for is a successful call that returns a low-confidence result you then send again, which is billed twice. Budget a few percent of your page volume for reprocessing, and more if you are sending crumpled thermal receipts or phone photos taken at an angle. How large that reprocessing share turns out to be depends on the engine, which is the other half of this decision: our breakdown of OCR accuracy covers what each vendor actually publishes and how to measure it on your own documents before you commit to a per page rate.

Build on an OCR API or buy a finished tool?

Run the math at your real volume. At 2,500 pages a month, Textract AnalyzeExpense costs $25 in API fees, which looks unbeatable next to a $49 plan until you add the pipeline someone has to build and keep running. At 500,000 pages a month, the API fee is $5,000 and a flat plan stops making sense, so building is clearly right. The crossover for most US teams lands somewhere in the tens of thousands of pages a month, and it moves depending on whether you already employ someone who can own the integration. Our cost comparison against manual data entry covers the third option, which is still what most small firms actually do.

What about per-seat expense platforms?

There is a fourth pricing model this table does not cover, because it is not billed per page at all. Expense platforms charge per person: Expensify publishes Collect at $5 per member per month and Control from $9 per active member, and Ramp lists Plus at $15 per user per month on top of a free tier. That model is priced off headcount rather than document volume, so it wins for a company with an expense policy and reimbursement cycle and loses badly for a bookkeeper processing thousands of documents for clients who will never log in. Our Expensify pricing breakdown works through the seat math and the billing rules, and Ramp against Expensify compares the two largest vendors on that side of the market.

Which API should I use for receipts specifically?

For receipts, use an endpoint trained on receipts rather than a general OCR call. Textract AnalyzeExpense, the Google expense parser, and the Azure prebuilt receipt model all return merchant, date, total, and tax as named fields, which is the work you would otherwise write yourself. Our comparison of Google Vision OCR against Document AI covers why the cheaper general OCR endpoint is usually a false economy for receipts, AWS Textract against Google Vision shows how identical headline rates buy very different output, and the Textract alternative page covers the same trade-off on the AWS side. For the per vendor rate cards behind this table, see Google Document AI pricing, Azure Document Intelligence pricing and the Google Cloud Vision API pricing breakdown.

OCR API pricing changes, so check the date

Vendors reprice, rename products, and move pricing behind sales forms. Google renamed its receipt parser to the expense parser. Azure renamed Form Recognizer to Document Intelligence. Klippa took its public pricing page down entirely, so any article quoting a specific Klippa number invented it. Dext moved the other way: it now publishes a full US rate card, at $25.21 a month for 250 documents and 5 users on its business plan, with line item extraction sold separately as an add-on from $20.50 a month. Every figure on this page carries the date it was checked, and figures we could not verify from a vendor source are marked as not published rather than filled in with a plausible guess. The same verified figures sit alongside the desktop and self-hosted options in our best OCR software comparison, and the licensing side of the free options is covered in our guide to open source OCR engines.