Research-backed calculator guide

Document AI Layout Extraction Cost: Price Text, Structure and Chunks Separately

Estimate document AI layout extraction cost by separating OCR, structure, chunking, retries, review, and storage instead of multiplying one headline page rate.

monthly pipeline cost
= successful layout pages × layout rate
+ successful OCR-only pages × OCR rate
+ re-chunked units × re-chunk rate
+ retry pages × retry rate
+ human review minutes ÷ 60 × loaded hourly rate
+ downstream storage, embedding, and generation cost

This is an estimate, not a vendor quote. Rates vary by provider, region, tier, contract, currency, processor version, and whether a feature is bundled. On November 2, 2026, the official Google Cloud rate card lists Layout Parser at $10 per 1,000 units and says it includes initial chunking; it separately lists re-chunking parsed documents at $0.02 per 1,000 units. Google also defines a page differently by file type, including one PDF page, one image, one spreadsheet tab, or up to 3,000 characters for some formats. ^[1] That distinction matters when the source workload is DOCX, HTML, or a parsed document rather than a PDF.

The shortest answer to “what will my layout extraction cost?” is: multiply the billable unit for the selected processor by the provider’s rate, then add operational steps that the headline price excludes. Do not call OCR, layout analysis, semantic chunking, embeddings, and human correction the same service.

What you are actually paying for

Text extraction produces searchable text. Google lists Enterprise Document OCR at $1.50 per 1,000 units for the first 5 million units on its displayed pay-as-you-go tier. Azure lists Read at $1.50 per 1,000 pages for the first million pages and $0.60 thereafter. ^{[1] [3]} These are useful baselines when you need text but not hierarchy.

Layout extraction detects relationships such as headings, paragraphs, tables, lists, headers, footers, and figures. Google’s layout-parser documentation describes a tree-like document layout and context-aware chunks that can retain ancestral headings and table context. ^[2] Azure describes its Layout option as detecting text and structure such as tables, titles, paragraphs, and selection marks. ^[3] Layout analysis can therefore be more valuable—and more expensive—than plain OCR even when both are measured in pages.

Chunking is a separate modeling choice. A provider may include an initial chunking pass in the layout processor, as Google does for its listed Layout Parser rate, while charging separately for re-chunking. ^[1] If your team only needs the provider’s default chunks, do not add a second chunk meter. If you change chunk size, overlap, metadata, or parent-heading behavior, record that as a separate stage.

Retries and review are pipeline costs, not always visible on a rate card. A retry can happen after a timeout, a low-confidence result, a malformed response, or a post-processing validation failure. Human review is usually best modeled from minutes per page or per document and a loaded hourly rate. Keep it outside API cost so a cheaper processor does not look cheaper merely because its quality creates more correction work.

Downstream retrieval is another boundary. Layout output prepared for RAG can feed embeddings, a vector store, and an LLM. Those costs belong in the complete document workflow, but they are not the same as layout extraction. Use the RAG, embedding and vector database cost calculator when the extracted chunks will be indexed.

Provider meters and comparison cautions

Google’s public rate card makes the separation unusually explicit: OCR, structure/entity processors, Layout Parser with initial chunking, and re-chunking appear as different rows. ^[1] Its rate card also says failed 4xx or 5xx requests are not billed, but that does not mean every application retry is free; confirm how your request and failure mode maps to the provider’s billing rules before assuming zero retry cost.

Azure’s public page lists $10 per 1,000 pages for all prebuilt models including Layout, $30 per 1,000 pages for custom extraction, and $1.50 per 1,000 pages for Read in the displayed US-dollar pay-as-you-go table. It says prices are estimates rather than quotes and that actual pricing can vary with agreement, purchase date, and exchange rate. ^[3] Azure also explains that billing is by analyzed pages and that all pages are analyzed unless a page range is specified. ^[3] For a fair comparison, use the same page range and do not compare an Azure Layout row with a Google OCR row as though they produce equivalent output.

AWS Textract presents another structure. Detect Document Text is a text-only API, while Analyze Document can add Forms, Tables, Queries, or Signatures and includes OCR in the output. Its published examples use region-specific per-page rates and show that multiple features can stack on the same pages. ^[4] The AWS page also gives a layout-and-tables example where layout is free when used with Tables, so a standalone “layout price” comparison can misstate the bill. ^[4]

The observed search results reflect this fragmentation. The exact query returned official pricing pages and layout documentation, a third-party comparison table, forum and Reddit billing questions, and an Azure-focused discussion. Most result formats answer one of three intents: “find a rate card,” “understand what the parser returns,” or “diagnose an unexpectedly high invoice.” The pages I inspected did not provide a neutral input worksheet for a multi-step layout-plus-chunking workflow with retries and human review. That is the useful coverage opportunity here—not a claim that any competitor is absent or that ranking or traffic is guaranteed.

Worked cost example

Start by fixing the accounting period and unit. Let D be documents per month, P average pages per document, and U = D × P submitted pages. Then estimate each stage. If only a fraction s of pages needs layout, layout pages are U × s; if the remaining pages use OCR-only, OCR pages are U × (1 - s).

The following table is an original example, not a vendor quote. It uses Google’s displayed list rates for illustration and assumes USD, pay-as-you-go pricing, no free allowance, no committed-use discount, and no taxes.

The example deliberately separates “API-only” from “pipeline” cost. API-only spend is $542.64: $480 for layout, $48 for OCR-only pages, $0.24 for re-chunking, and $14.40 for retries. Review adds $180. The modeled cost per accepted document is therefore about 7.85 cents, while API-only cost per accepted document is about 5.90 cents. If you report only the $480 headline layout line, you hide 33.6% of the modeled pipeline cost in this example.

Three assumptions deserve explicit testing. First, “60% require layout” should come from a sample of real documents, not a guess. Second, the 3% retry rate should be measured by reason and provider response; a timeout retry is different from a quality-driven second pass. Third, 92% acceptance is a business-quality threshold, not a provider promise. Change it in a sensitivity table before committing to a budget.

For example, holding all other inputs constant, raising review from 6 to 10 minutes per reviewed document adds $120. Raising layout coverage from 60% to 80% shifts 16,000 pages from the $1.50 OCR assumption to the $10 layout assumption, adding $136 before retry effects. Those two variables can matter more than the $0.24 re-chunking row. This is why a price-per-page comparison alone is not a cost model.

How to budget retries, review, and RAG

Use a retry reserve only for attempts your billing evidence supports. A conservative worksheet has separate inputs for transport retries, provider errors, low-confidence reprocessing, and human-requested corrections. If the provider does not charge for a failed request, set that specific row to zero while keeping successful second passes in the paid row. Do not use a generic “10% contingency” without labeling what it represents.

For review, measure queue minutes per document and the fraction routed to a person. The cost is reviewed documents × minutes per document ÷ 60 × loaded hourly rate. Include wages, payroll burden, management, and tooling in the loaded rate if the result is meant for unit economics. If review is outsourced, replace the estimate with the contracted rate and date it.

For RAG, export page and block identifiers with each chunk. That makes it possible to trace a bad answer to the layout pass, chunking rule, embedding record, or generation step. Use the AI API invoice reconciliation calculator to compare these modeled stages with invoice lines after launch. A provider’s API rate is a forecast input; your invoice is the observed result.

Do not forget file-type and request-limit assumptions. Google’s layout-parser documentation lists different limits for online and batch processing and notes that supported file types are subject to charges. ^[2] It also warns that tables spanning multiple PDF pages can be split and that nested tables or hidden spreadsheet content can affect detection. ^[2] A lower price is not a saving if the output requires a costly repair pass.

When comparing providers, normalize at least five things: pages analyzed, feature set, region, monthly volume tier, and output acceptance rule. Then record whether chunking is included, whether retries are billable, whether a free tier applies, and whether there are hosting or commitment charges. The model pricing comparison guide is useful for keeping unit, tier, and date assumptions visible across vendors.

FAQ

Is layout extraction more expensive than OCR?

Often, but not always. Plain OCR and layout extraction are different outputs, and providers price them separately or bundle features differently. Google’s displayed rates put Enterprise Document OCR at $1.50 per 1,000 units and Layout Parser at $10 per 1,000 units, while Azure lists Read at $1.50 and Layout among prebuilt models at $10 per 1,000 pages. ^{[1] [3]} AWS can bundle layout with Tables, so compare the requested feature set rather than the word “layout.” ^[4]

Is document AI billed per document or per page?

Many document AI meters are page-based, but the definition of a page varies. Google says one PDF page or image counts as one page, while a DOCX, HTML, or parsed document can use a character-based definition. ^[1] Azure says it bills analyzed pages and analyzes all pages unless a range is specified. ^[3] Record the provider’s unit and your file mix before multiplying by document count.

Does initial chunking cost extra after layout parsing?

It depends on the processor. Google’s listed Layout Parser price includes initial chunking, while re-chunking parsed documents has a separate displayed rate. ^[1] If you accept the provider’s initial chunks, do not add a second initial-chunk charge. If you rerun chunking with a different strategy, model that stage explicitly.

Should human review be included in layout extraction cost?

Include it when the goal is cost per accepted document, but show it separately from API-only cost. Review depends on routing rate, minutes per item, and loaded labor cost. A layout processor that costs less per page can still have a higher accepted-output cost if it routes more documents to review.

What should I record for a production estimate?

Record document count, pages per document, file types, layout coverage, OCR-only coverage, retry rate and reason, chunking mode, provider region, price tier, review rate, review minutes, and acceptance rate. Add storage, embeddings, and generation only when they belong to the costing boundary. Date every vendor rate and mark estimates separately from quotes.