amBrain

AWS Textract alternative: rent another OCR API or own the document pipeline

Published Sep 23, 2026Facts checked: Sep 23, 2026

Short answer

Stay on Amazon Textract if your documents are in English or one of the five other languages its printed text, forms and tables cover, your stack already runs on AWS, and per-page prices of $0.0015 for text detection or $0.05 for forms fit your volume.[1][2] Look elsewhere if nobody can show that the AWS Organizations opt-out was set, so your documents may be used to improve Amazon's models, if you need a region outside Textract's 15, or if the forms-and-tables bill grows with every page.[2][1] The alternatives compared here are Google Document AI, Azure Document Intelligence with its containers, and your own pipeline on open-source OCR and a self-hosted model.

On this page

What Textract charges, and what it does with your documents

Textract prices per page and per API, and AWS publishes both the rates and worked examples, so a bill can be calculated before anything is built.

In US West (Oregon), text detection costs $0.0015 a page for the first million pages a month and $0.0006 after that; tables $0.015 then $0.01; forms $0.05 then $0.04; Queries $0.015; Custom Queries $0.025 then $0.015; invoices and receipts (Analyze Expense) $0.01 then $0.008; identity documents $0.025 for the first 100,000 pages then $0.01; lending documents $0.07 then $0.055. AWS's own worked examples put 100,000 pages of plain text at "$150" a month and two million pages with forms and tables at "$115,000" a month, and the free tier is time-boxed: "The Free Tier lasts for three months".[1]

The data terms are in the FAQ, and they have a default. "You always retain ownership of your content, and we will only use your content with your consent", yet Textract "may store and use document and image inputs … to improve and develop the quality of Amazon Textract and other Amazon machine-learning/artificial-intelligence technologies", and "Unless you opt out as provided below, some portion of content processed by Amazon Textract may be stored in another AWS region…". The opt-out is a policy somebody has to set ("You may opt out … using an AWS Organizations opt-out policy."), and deletion is a request: "You can request deletion of document and image inputs associated with your account by contacting AWS Support."[2]

Scope has limits too. Textract runs in 15 AWS regions: N. Virginia, Ohio, Oregon, N. California, GovCloud US-West, GovCloud US-East, Canada Central, Ireland, London, Frankfurt, Paris, Singapore, Sydney, Seoul and Mumbai. Printed text, forms and tables cover English, German, French, Spanish, Italian and Portuguese, but the same page narrows the rest: "Handwriting, Invoices and Receipts, Identity documents and Queries processing are in English only." Handwriting detection covers "the Standard English alphabet and ASCII symbols".[2]

Textract and the alternatives compared

Four options, by what each vendor publishes, read on 2026-09-23
DimensionAmazon TextractGoogle Document AIAzure Document IntelligenceOwn pipeline (open-source OCR + self-hosted model)
Code and data ownership"You always retain ownership of your content", but inputs "may" be stored and used to improve Amazon AI unless an AWS Organizations opt-out policy is set[2]"At Google Cloud, we never use customer data to train our Document AI models." "For batch operations, the stored document is typically deleted immediately after the processing, with a failsafe Time to live (TTL) of one day."[3]"Document Intelligence temporarily stores data and results in Azure Storage in the same region as the request. Your data is then deleted 24 hours from the time that you submit an analyze request."[6]Documents never leave servers you operate; the code is yours
Cost modelPer page in US West (Oregon), first million pages a month: $0.0015 text detection, $0.015 tables, $0.05 forms, $0.015 Queries[1]Per 1,000 pages: Enterprise Document OCR free to 1,000, $1.50 from 1,000 to 5,000,000 and $0.60 above; Form Parser $30.00, $20.00 above a million; custom extractor $30.00 then $20.00; Layout Parser "(Includes initial chunking) $10.00"; custom splitter and custom classifier $5.00 then $3.00[4]Per 1,000 pages on S0: Read $1.50, $0.60 above 1,000,000 pages; pre-built $10.00; custom $30.00; training $3.00 an hour[9]Servers, GPUs and people; no per-page fee
Time to launchAn AWS account and an API callA Google Cloud project and a processorAn Azure resource, or a container you runScoped per project
Lock-in: term, notice, deprecation clockNo commitment terms on the pricing page, which does carry "For high volume use cases, connect with our sales team to request a custom pricing proposal."[1]Pay per page in monthly volume tiers; the same page also sells a "Gemini Enterprise Flexible Savings Plan — 1 Year" and "— 3 Year" consumption model, with a discounted column on every table, and a "Service tiers … Provisioned" reservation at $0.410958904 an hour[4]REST API v2.1 end of support 15 September 2027, v3.0 (2022-08-31) 30 March 2029[7]; monthly commitment tiers are sold alongside pay-as-you-go[9]None; you choose when to change a model
CustomisationQueries ("For synchronous operations, a maximum of 15 Queries per page is supported. For asynchronous operations, a maximum of 30 queries per page is supported.") and Custom Queries adapters. Printed text, forms and tables cover six languages, while "Handwriting, Invoices and Receipts, Identity documents and Queries processing are in English only."[2]Custom extractor, custom classifier and custom splitter, each with its own per-1,000-page price[4]Custom extraction and classification models at $30.00 per 1,000 pages, with training at $3.00 an hour[9]Any OCR engine, model, schema and validation rule
Compliance and data residency15 AWS regions; HIPAA-eligible with an executed Business Associate Agreement[2]Multi-region us and eu, plus the single regions Mumbai, Singapore, Sydney, London, Frankfurt, Amsterdam and Montréal[5]Cloud or containers: "Container pricing is the same as cloud service pricing"[8]; disconnected containers are billed by the year, the "Pre-Built Disconnected 100K" meter at $8,640.00[9]Wherever your servers are; you produce the evidence
Latency and scaleService quotas per account, and AWS advises smoothing the load: "Spiky traffic affects throughput…'smooth' traffic"[2]Volume tiers: Enterprise Document OCR drops from $1.50 to $0.60 per 1,000 pages above five million a month[4]Commitment tiers from 500,000 Read pages a month at $375.00 to 16,000,000 at $7,200.00[9]; the pricing page lists a free tier of 500 pages a month[8]Set by your hardware and your queue design
Exit costRe-validate every field on the new engine; neither the pricing page nor the FAQ says whether a Custom Queries adapter can be exported or moved[1][2]Re-validate every field on the new engineRe-validate, and migrate API versions before the end-of-support dates[7]None contractual

Stay on Amazon Textract if…

  1. 1.Your documents fit its scope. Printed text, forms and tables in English, German, French, Spanish, Italian or Portuguese, with handwriting, invoices and receipts, identity documents and Queries "in English only",[2] stored in S3 in one of the 15 regions.
  2. 2.Your volume keeps the bill small. AWS's own example puts 100,000 pages of plain text at "$150" a month;[1] at that bill your own pipeline cannot pay back the engineering time it needs.
  3. 3.You need a HIPAA-eligible managed service. Textract is HIPAA-eligible with an executed Business Associate Agreement,[2] and you do not want to operate OCR infrastructure.

Consider leaving Amazon Textract if…

  1. 1.Symptom: nobody can prove the opt-out was set. A customer's security questionnaire asks whether documents are used for model training, and the honest answer depends on an AWS Organizations policy no one has checked. Without it, inputs may be used to improve Amazon's AI and "some portion of content processed by Amazon Textract may be stored in another AWS region…".[2] The same question for LLM APIs on the same account is in AWS Bedrock alternatives.
  2. 2.Symptom: forms and tables drive the bill. The invoice tracks page count, not value: AWS's own example puts two million pages with forms and tables at "$115,000" a month.[1] Google lists Form Parser at $30.00 per 1,000 pages,[4] while Textract's published $0.05 a page for forms is $50 per 1,000 by arithmetic.[1]
  3. 3.Symptom: your documents are outside its languages. A manual correction queue grows with every new market, because printed-text extraction covers six languages, handwriting covers "the Standard English alphabet and ASCII symbols", and "Handwriting, Invoices and Receipts, Identity documents and Queries processing are in English only."[2]

A migration path in four steps

  1. 1.Build a labelled acceptance set. Per document type and per field, with the values a person accepts. It lets you compare engines on your documents rather than on demos; the AI pilot that never reached production explains why it comes before any model choice.
  2. 2.Run the candidate in shadow. Send the same pages to Textract and to the candidate, whether Google, Azure or your own pipeline, and score each field against the acceptance set.
  3. 3.Switch by document type. Move the types where the candidate passes; keep Textract for the rest until they pass too. How an in-house pipeline is built stage by stage, from classification and extraction to validation, review and audit, is in an LLM pipeline for claims, tickets and KYC files.
  4. 4.Close the AWS side. Set the AI services opt-out policy if you have not, ask AWS Support to delete stored inputs,[2] and remove the Textract permissions from your roles.

What drives the cost of building your own

These are the cost drivers, not a price.

Where the money goes in a pipeline you run, and what can be priced from public pages
Cost driverWhat drives itPublic reference point
OCR engineLanguages, layouts, scan qualityOpen-source engines, none of the four repositories archived: Tesseract (Apache-2.0)[10], PaddleOCR (Apache-2.0)[11], docTR (Apache-2.0)[12], Docling (MIT)[13]
Extraction modelField complexity, free text versus fixed formsA self-hosted open-weight model: gpt-oss-20b runs "within 16GB of memory" and gpt-oss-120b "on a single 80GB GPU (like NVIDIA H100 or AMD MI300X)"[14]
HardwarePages per hour at peak, GPU or CPU per stageCompare with the rate you replace: Textract forms at $0.05 a page[1]
Review queueShare of fields a person must confirmNo public reference; measured on your acceptance set
Evaluation and labellingLabelled pages per document type, rerun on every changeNo public reference; scoped per project
Audit trailWhich model and rule produced which fieldNo public reference; part of the pipeline design
OperationsQueues, retries, monitoring, on-callNo public reference; scoped per project

Two decisions sit inside that table. Whether the extraction model is rented or run on hardware you control is weighed in OpenAI API vs a self-hosted LLM; which server runs it, in vLLM vs Ollama. If the model stays rented, OpenAI API vs Claude API compares the two published sets of terms.

Who builds a document pipeline that stays inside your infrastructure

Look for a team that asks for your documents and your acceptance criteria before it names a model.

About amBrain

  • We have taken an LLM integration to production inside a client's FinTech perimeter: extracting and normalising unstructured broker and venue notices — corporate actions, instrument and margin changes — into structured records the trading system consumes.
  • The client keeps full ownership of the product and the code, except our reusable components.

Frequently asked questions

It may, unless you opt out. The FAQ says Textract "may store and use document and image inputs … to improve and develop the quality of Amazon Textract and other Amazon machine-learning/artificial-intelligence technologies", that "You may opt out … using an AWS Organizations opt-out policy.", and that "You can request deletion of document and image inputs associated with your account by contacting AWS Support."[2]

In US West (Oregon), for the first million pages a month: $0.0015 for text detection, $0.015 for tables, $0.05 for forms, $0.015 for Queries, $0.025 for Custom Queries and $0.01 for invoices and receipts, with lower rates above a million pages. New customers get a time-boxed free tier: "The Free Tier lasts for three months".[1]

Printed text, forms and tables cover English, German, French, Spanish, Italian and Portuguese, and handwriting detection covers "the Standard English alphabet and ASCII symbols". The rest is narrower than the list suggests: "Handwriting, Invoices and Receipts, Identity documents and Queries processing are in English only."[2]

For plain OCR the published rates line up: Google's Enterprise Document OCR is $1.50 per 1,000 pages after a free first 1,000,[4] and Textract's published $0.0015 a page is $1.50 per 1,000 by arithmetic.[1] For forms, Google lists Form Parser at $30.00 per 1,000 pages[4] against Textract's $0.05 a page, $50 per 1,000 by the same arithmetic.[1] Score both on a sample of your own documents before choosing by list price.

Azure Document Intelligence runs in containers, and "Container pricing is the same as cloud service pricing",[8] including disconnected containers billed by the year: the "Pre-Built Disconnected 100K" meter is $8,640.00.[9] Or build your own pipeline from open-source OCR (Tesseract,[10] PaddleOCR,[11] docTR[12] or Docling[13]) and a self-hosted open-weight model such as gpt-oss-20b, which runs "within 16GB of memory".[14]

Disclosure: this page is published by amBrain. The last column of the comparison table is the kind of work amBrain does. Every vendor fact on this page links to the page it was read from on 2026-09-23; document-AI prices change often, and figures given per 1,000 pages for Textract are arithmetic on AWS's published per-page prices. No vendor named here reviewed or endorsed this page. Amazon Textract, Amazon Web Services, Google Cloud, Google Document AI, Microsoft, Azure AI Document Intelligence, Tesseract, PaddleOCR, docTR and Docling are trademarks of their owners and are used only to identify the products discussed.

Sources

  1. [1]Amazon Web Services, Amazon Textract pricing (per-page prices by API in US West (Oregon), the worked examples, the free tier and the custom-pricing line). Vendor's own page · checked Sep 23, 2026
  2. [2]Amazon Web Services, Amazon Textract FAQs (data use and the opt-out, deletion requests, the region list, languages, handwriting, Queries limits, HIPAA eligibility and throughput advice). Vendor's own page · checked Sep 23, 2026
  3. [3]Google Cloud, Document AI security and compliance (customer-data usage and how long a processed document is stored). Official documentation · checked Sep 23, 2026
  4. [4]Google Cloud, Document AI pricing (per-1,000-page tables by processor, the Flexible Savings Plan columns and the provisioned service tier). Vendor's own page · checked Sep 23, 2026
  5. [5]Google Cloud, Regional and multi-regional support. Official documentation · checked Sep 23, 2026
  6. [6]Microsoft, Document Intelligence FAQ (where analysis data is stored and when it is deleted). Official documentation · checked Sep 23, 2026
  7. [7]Microsoft, What's new in Document Intelligence (end-of-support dates for the REST API versions). Official documentation · checked Sep 23, 2026
  8. [8]Microsoft, Azure AI Document Intelligence pricing (the free tier of 500 pages a month and the container-pricing line). Vendor's own page · checked Sep 23, 2026
  9. [9]Microsoft, Azure Retail Prices API, Document Intelligence meters in Central US (S0 per-1,000-page meters, training, commitment tiers and the disconnected-container meter). Vendor's own page · checked Sep 23, 2026
  10. [10]Tesseract OCR, Source repository and licence. Vendor's own page · checked Sep 23, 2026
  11. [11]PaddlePaddle, PaddleOCR source repository and licence. Vendor's own page · checked Sep 23, 2026
  12. [12]Mindee, docTR source repository and licence. Vendor's own page · checked Sep 23, 2026
  13. [13]Docling project, Docling source repository and licence. Vendor's own page · checked Sep 23, 2026
  14. [14]OpenAI, gpt-oss-120b model card (memory needed to run each of the two open-weight sizes). Vendor's own page · checked Sep 23, 2026

Related

Free project intro call, 30 minutes.

We reply within 24 hours.

Book a Call