Direct answer: public document-OCR prices can look very low in 2026: AWS Textract charges USD 0.0015 per page for plain text in the first million pages, and Google Document AI Enterprise OCR charges USD 1.50 per 1,000 pages for the first five million. But tables, forms, invoices, and custom fields move the price to USD 0.015–0.10 per page or document. At 100,000 pages/month, plain text is about USD 150, tables about USD 1,500–3,000, and forms up to USD 5,000; human review, integration, and exception handling often cost more than the OCR API. Alibaba Cloud OCR uses tiered per-call pricing; its ID-card example is RMB 0.0825/0.0495/0.0415 per call. This is a same-basis cost study, not a vendor ranking.
This guide is for manufacturers, finance shared-service teams, IT/data teams, and procurement teams digitising invoices, inspection reports, COAs, packing lists, drawings, or archives. Data cut-off: 31 October 2026. Prices come from official AWS, Google Cloud, Alibaba Cloud, and Microsoft Azure pages. All prices are public list prices, excluding tax; enterprise agreements, regional differences, commitments, and invoices control. This is an editorial model, not a quote, and does not promise accuracy, savings, or a go-live date.
Plain-text OCR is cheap; tables, forms, and invoices cost an order of magnitude more. Per page, per document, and per call are different billing units. Human review and exception handling are usually the largest TCO line.
What are the official OCR billing units in 2026?
AWS and Google mainly bill per page; Google’s invoice parser bills per document; Alibaba Cloud bills per API call on a tiered schedule. Align the unit before comparing price.
“OCR” is not one product. Plain text, table extraction, form fields, invoice/expense parsing, ID recognition, and custom extraction have different prices and units. The table lists only public list prices verified during this check and does not guess dynamic or sales-quoted prices.
| Service / capability | Billing unit | Official list price | Note |
|---|---|---|---|
| AWS Textract plain-text OCR | Page | USD 0.0015/page (first 1M); USD 0.0006/page (above) | Detect Document Text API; US West Oregon example |
| AWS Textract tables | Page | USD 0.015/page (first 1M); USD 0.010/page (above) | Analyze Document Tables; Layout is free with Tables |
| AWS Textract forms | Page | USD 0.05/page (first 1M); USD 0.04/page (above) | Forms key-value pairs; add-on charges stack |
| AWS Textract expense | Page | USD 0.01/page (first 1M); USD 0.008/page (above) | Analyze Expense API |
| Google Document AI Enterprise OCR | Page | USD 1.50 per 1,000 pages (1,000–5M); USD 0.60 per 1,000 above | First 1,000 pages/month free |
| Google Document AI Form Parser | Page | USD 30 per 1,000 pages (first 1M); USD 20 per 1,000 above | Form/custom extraction; not plain OCR |
| Google Document AI Invoice Parser | Document (≤10 pages) | USD 0.10/document | One count covers up to 10 pages |
| Alibaba Cloud OCR ID card | Call | RMB 0.0825 (≤10k); RMB 0.0495 (10k–100k); RMB 0.0415 (100k–500k) | ID-card example; different APIs differ; 200 free calls/API/month |
Official list prices, USD and RMB listed separately, excluding tax, as of 31 October 2026. AWS uses US West (Oregon) examples; Google is USD list price; Alibaba is tiered pay-as-you-go, with cheaper one-year resource packs. Azure Document Intelligence also bills per page with 500 free pages/month, but its price page rendered values dynamically and no stable unit price was captured, so it is not in the numeric comparison.
What do 10k, 100k, and 1M pages cost per month?
Plain text is about USD 15/150/1,500; tables about USD 150/1,500/15,000; forms up to USD 500/5,000/50,000. Units differ, so do not compare one row.
| Option | 10,000 pages/month | 100,000 pages/month | 1,000,000 pages/month | Unit |
|---|---|---|---|---|
| AWS plain-text OCR | USD 15 | USD 150 | USD 1,500 | Page |
| AWS tables | USD 150 | USD 1,500 | USD 15,000 | Page |
| AWS forms | USD 500 | USD 5,000 | USD 50,000 | Page |
| AWS expense | USD 100 | USD 1,000 | USD 10,000 | Page |
| Google Enterprise OCR | USD 15 | USD 150 | USD 1,500 | Page |
| Google Form Parser | USD 300 | USD 3,000 | USD 30,000 | Page |
| Google Invoice Parser | USD 1,000 | USD 10,000 | USD 100,000 | Document (≤10 pages) |
Editorial model: pages × unit price at the first tier; excludes tax, human review, integration, storage, retries, and exception handling. Google Invoice Parser is modelled as 10,000/100,000/1,000,000 documents, each ≤10 pages; it is not directly additive or comparable with per-page prices. Tier discounts above the first band are not included.
Why is “USD 0.0015 per page” not the bill?
The second bill sits in table/form/invoice capability, human review, integration, retries, and data compliance. Plain-text OCR is only the entry point.
Document intelligence has at least seven cost lines. Plain-text OCR is cheapest but returns characters only. To get fields, tables, invoice number/amount/tax, you need form, table, invoice, or custom models, often 10–30× the plain-text price. Drawings, handwriting, stamps, skew, and cross-page tables raise failure and review rates.
| Cost line | Included in API price? | How to meter | Common omission |
|---|---|---|---|
| Page/document/call fee | Yes | Pages × price; documents × price; calls × tier | Applying a per-page price to a per-document or per-call API |
| Custom training/hosting | Usually not | Training hours, hosting hours, model versions | Counting inference only |
| Human review/exceptions | No | Exception rate × minutes/page × labour rate | Assuming 100% auto-pass |
| Integration/process change | No | APIs, queues, retries, ERP/finance write-back | OCR without reconciliation |
| Storage/security/compliance | Partly | Retention, region, encryption, permissions, audit | Sending client/PII data to an unapproved region |
| Failures and reruns | No | Failure rate, retries, duplicate charges | Budgeting only successful pages |
| Maintenance/versioning | No | Template/field changes, upgrades, regression | No owner after go-live |
Editorial guidance, not an accounting taxonomy. Follow corporate finance policy, vendor contracts, and actual invoices.
At 100k pages/month, how much is human review?
If 2% of pages need review, at 5 minutes/page and RMB 60/hour, review is about 166.7 hours and RMB 10,000 — potentially more than the plain-text OCR API fee.
| Exception-rate assumption | Pages to review | Review hours (5 min/page) | Labour cost (RMB 60/hour) | Note |
|---|---|---|---|---|
| 1% | 1,000 | 83.3 hours | RMB 5,000 | Stable layout, few fields |
| 2% | 2,000 | 166.7 hours | RMB 10,000 | Medium complexity |
| 5% | 5,000 | 416.7 hours | RMB 25,000 | Handwriting/stamps/cross-page tables |
100,000 pages/month. Exception rate, 5 minutes/page, and RMB 60/hour are editorial assumptions, not an industry standard or quote. API fees are in USD and labour in RMB; they are listed separately and not converted or added. Replace them with your own rates.
This explains why “USD 150 per 100k plain-text pages” is not a project budget: at a 2% exception rate, human review can be RMB 10,000, far above the API fee; cutting exceptions to 1% halves the review cost. The optimisation order is standardise layouts → add field/table models → route by confidence → review only exceptions, not simply switching to a cheaper OCR engine.
How to choose by scenario
For plain-text archives, choose the cheapest OCR; for tables/forms, choose structured extraction; for invoices, compare per document; for drawings/handwriting, run a small acceptance set instead of using the plain-text price.
| Scenario | Start with | Why | Do not ignore |
|---|---|---|---|
| Contract/archive full-text search | Plain-text OCR | Lowest unit price; simple per-page billing | Character accuracy, page order, language |
| Inspection reports/COA/packing lists | Tables + Layout | Need row/column and cell relationships | Cell accuracy, cross-page and merged cells |
| Invoices/expense claims | Invoice/Expense Parser | Per document or per page; fields are amount/tax/date | Unit, 10-page limit, credit notes |
| Drawings/handwriting/stamps | Custom extraction + review | Generic OCR cannot guarantee fields/layout | Training samples, exception rate, review hours |
| China data-boundary requirements | Regional cloud/private/contract | Residency, permissions, audit can be confirmed | Do not compare price alone; require DPA/security review |
| Above 1M pages/month | Commitment/tiered pricing | Unit price falls after the first band; negotiate commitment | Minimum spend, overage, lock-in |
Editorial guidance, not a vendor ranking. Document types, layout stability, language, scan quality, and compliance differ; run a 1,000–5,000 page pilot and measure accuracy and review rate on the same field set.
What should procurement and acceptance measure?
Measure field accuracy, table-cell accuracy, exception review rate, cost per 1,000 pages, and data boundaries — not character accuracy or page price alone.
| Acceptance item | Suggested definition | Evidence |
|---|---|---|
| Field accuracy | By document type and field; not one aggregate score | Labelled sample, rubric, confusion cases |
| Character accuracy | CER/WER or human sampling; Chinese/English/digits separate | Raw OCR text, sampling record |
| Table-cell accuracy | Row/column/cell mapping, including merged cells | Structured output, manual check |
| Exception review rate | Pages requiring human handling ÷ total pages | Confidence threshold, review queue |
| Cost per 1,000 pages | (API + training/hosting + review + integration) ÷ 1,000 pages | Invoice, hours, project input |
| Failure/retry | Failed requests, rerun pages, duplicate charges | API logs, billing detail |
| Data boundary | Region, retention, encryption, permissions, audit, DPA | Contract, configuration, audit logs |
| Version/maintenance | Template/model version, regression set, change log | Version record, regression report |
There is no universal “good enough” threshold; invoices, COAs, and drawings carry different field risks. Define fields that must be 100% correct (amount, tax, batch number), then set sampling/review rules for the rest.
Who should read this — and who should not?
For finance/IT/data teams digitising invoices, inspection reports, archives, or drawings. Not for a few dozen pages, no layout samples, or a project that refuses human review.
Read if: you work in finance shared services, procurement, IT/data, archives, quality, or manufacturing engineering; you are comparing AWS Textract, Google Document AI, Alibaba Cloud OCR, or Azure Document Intelligence; you need to compare per page/document/call, estimate human review, and define acceptance. Skip if: you have only a few dozen pages and expect free tiers to solve production; you have no labelled samples or layout inventory; you expect 100% OCR accuracy with no exception handling; you plan to upload client PII or drawings to an unapproved region. Supports: unit conversion, 10k/100k/1M-page models, review TCO, scenario selection, and acceptance metrics.
Lingyi does not rank OCR vendors or promise accuracy. If documents cannot leave the intranet and you need private deployment, permissions, audit, or multi-system integration, see LingHarness; for a lightweight trial of public OCR APIs, the cloud vendor’s official service is usually faster. To scope document types, fields, and acceptance, talk to us.
See our separate notes on RAG knowledge-base cost, private-deployment TCO, and AI project RFPs; this article only covers OCR/document-intelligence page and call costs.
FAQ
Why is the actual bill higher than USD 0.0015 per page?
USD 0.0015 is the first-tier plain-text OCR price. Tables, forms, invoices, custom extraction, and ID recognition are different APIs, often 10–30× the plain-text price; add human review, integration, storage, retries, and compliance. Confirm the capability before comparing unit price.
Are tables billed per page or per table?
AWS Textract and Google Document AI table/form capabilities bill per page; a cross-page table still consumes pages. Google’s Invoice Parser bills per document, with one count covering up to 10 pages. Ask how “page” is defined, and whether cross-page tables and blank pages are charged.
Are invoices billed per page or per document?
It depends. AWS Analyze Expense bills per page; Google Invoice Parser bills USD 0.10 per document (≤10 pages). 100,000 one-page invoices and 100,000 multi-page invoices can cost very differently; compare on the same document/page basis.
Why does Alibaba Cloud OCR bill per call?
Alibaba Cloud OCR bills per successful API call on a tiered schedule; different OCR APIs have different prices. The ID-card example is RMB 0.0825/0.0495/0.0415 per call, with 200 free calls per API per month. It cannot be compared directly with per-page document AI until you convert actual calls per page/document.
What does 100,000 pages/month cost?
At public first-tier prices: plain-text OCR about USD 150; AWS tables about USD 1,500; AWS forms about USD 5,000; AWS expense about USD 1,000; Google Form Parser about USD 3,000; Google Invoice Parser about USD 10,000 for 100,000 documents of ≤10 pages. Add human review, integration, and compliance; the table is an editorial model, not a quote.
How should we accept OCR ROI?
Use the same 1,000–5,000 page sample, measure field accuracy by document type, table-cell accuracy, exception review rate, cost per 1,000 pages, and failures/retries; include human review hours. Comparing only page price or a vendor’s character-accuracy demo does not prove your layouts and fields will go live.
Method and limitations
This article uses only official pricing pages accessible on 31 October 2026. It does not use private contracts, customer invoices, or vendor internal data. Prices, regions, free tiers, first/overage bands, and billing units can change. The article has no actual enterprise agreement, tax, FX, human review, or integration data, so every monthly cost is an editorial model, not a quote. Azure Document Intelligence bills per page with 500 free pages/month, but its price values rendered dynamically and no stable unit price was captured, so it is excluded from the numeric comparison. Alibaba Cloud’s ID-card example does not represent every OCR API.
Sources and basis
Data cut-off: 31 October 2026. The following are official pricing/product pages; prices, regions, free tiers, and terms follow the original page and contract.
- AWS, Amazon Textract pricing: aws.amazon.com/textract/pricing. Per-page prices and examples for text, tables, forms, expense, and ID.
- Google Cloud, Document AI pricing: cloud.google.com/document-ai/pricing. Enterprise OCR, Form Parser, and Invoice Parser per-page/per-document prices and free tier.
- Alibaba Cloud, OCR billing: help.aliyun.com/zh/ocr/product-overview/billing. Pay-as-you-go tiers, resource packs, free quota, and ID-card billing example.
- Microsoft Azure, Document Intelligence pricing: azure.microsoft.com/pricing/details/ai-document-intelligence. Per-page billing and 500 free pages/month; dynamic values not captured.