Document Parsing API Costs, File by File
Compare seven document parsing APIs by output and billing unit, then price mixed-length files, page selection and a second pass without guessing quality.
Document parsing can cost $1.50 per 1,000 OCR pages with Amazon Textract or Google Document AI, $10 per 1,000 pages for Google’s Layout Parser, or $30 per 1,000 pages for its Form Parser at the listed first tiers. Those rates buy different outputs. A Google invoice parser instead lists $0.10 for each document block of up to ten pages, so 200 mixed-length files can cost more than a page-total estimate suggests. First choose the output, then count the units for each file and operation. These are published tariffs, not a quality ranking or a quote for your account. Amazon Textract pricing and Google Document AI pricing give the example rates.
Choose the output before comparing prices
Plain OCR returns words and their positions. Layout parsing identifies structure such as headings and tables; that can be useful before search or retrieval. Field extraction aims to return named values such as an invoice total. A rate for plain text cannot be treated as the price of reliable fields. Google’s extraction guide distinguishes Form Parser, Custom Extractor and Layout Parser. Textract’s pricing page separates Detect Document Text from Analyze Document features including Forms, Tables and Queries.
Write down the output you require before shopping: searchable text, table structure, retrieval chunks, or fields in a schema. Then specify whether your input is a scanned PDF, a digital PDF, images or mixed files. A vendor can offer several routes with different tariffs. The table below names one route per provider; changing routes changes the comparison.
Compare seven document parsing routes
USD list prices below were read on November 7, 2026. “Per 1,000 pages” describes a displayed rate, not necessarily a thousand-page minimum or the final bill. Free allocations and contracts can change an account’s net charge. The output column prevents a low OCR price from masquerading as a field-extraction price.
| Provider and named route | Published entry rate or limit | Billable unit and output |
|---|---|---|
| Amazon Textract Detect Document Text | $0.0015/page for the first million in the US West (Oregon) pricing example; new-customer free tier lists 1,000 pages/month for three months | Processed page; text OCR. Analyze Document Forms and Tables are separate feature prices. |
| Google Document AI Enterprise OCR / Layout Parser / Form Parser | $1.50 / $10 / $30 per 1,000 pages at the listed first paid tiers; OCR shows 1,000 free pages in its table | Processed page; respectively OCR, layout and chunks, or generic form fields. Its invoice parser is a different $0.10 per up-to-ten-page document block. |
| Azure Document Intelligence Read / Layout / Prebuilt | Public page shows 500 free pages/month; paid numbers rendered as “$-” in the readable table, so a paid rate is not established here | Pages analyzed; Read text, Layout structure, or prebuilt fields. The pages request parameter can select a range. |
| Mistral OCR OCR 4.1 | $4 per 1,000 pages in its standard pricing table | Page; OCR and structured document output. Check model and processing mode before using a batch or cached rate. |
| LlamaParse v2 Fast / Cost-effective / Agentic | 1 / 3 / 10 credits per page; Pay-as-you-go credits cost $1.50 per 1,000, making the listed parse charge $1.50 / $4.50 / $15 per 1,000 pages before included credits | Page multiplied by selected tier credits; Fast is for plain-text documents. Extraction and indexing are separate products. |
| Reducto r-1 Parse / Extract | Standard lists $10 / $20 per 1,000 pages; Extract includes parsing; $150 in free usage is advertised | Page and chosen endpoint; structured parse or field extraction. Growth and Enterprise are custom quoted. |
| Adobe PDF Extract | 500 free Document Transactions/month; paid plan price is not published on the public page | Document Transaction; PDF content and structural extraction. Ask Adobe for the paid transaction tariff and exact metering terms before comparing long files. |
The table does not imply that seven outputs are equivalent. Textract’s Oregon example is region-specific. Azure’s public table gives no usable paid number in the version read, and Adobe invites sales contact for paid plans. A blank should stay a blank rather than become a competitor’s estimate. LlamaParse’s credit cost is a calculation from its current developer documentation, so an included-credit plan can produce a different invoice. LlamaParse pricing describes the tiers and billing describes the credit purchase rate.
Count the units for every document
Build a worksheet with one row per file: file ID, physical pages, pages submitted, endpoint, billable units, retries, and any second operation. For page-priced OCR, calculate processed pages × the applicable per-page rate. If one API response includes the needed feature, do not automatically add another OCR call. Textract says Analyze Document includes OCR, while its Forms and Tables combination has feature pricing of its own. Textract pricing shows those combinations.
For a route that bills in blocks of B pages per document, calculate units = sum(ceil(pages_i / B)) across files. Do not use ceil(sum(pages_i) / B): that discards the unused capacity inside each file’s last block. Google’s invoice parser pricing explicitly defines one count as up to ten pages in a document and gives $0.10 for one to ten pages, $0.20 for eleven to twenty. That is the documented block case here; it does not describe Google’s Enterprise OCR or Form Parser tariffs.
Other metering needs its own formula. LlamaParse multiplies billed pages by tier credits and then applies the plan’s credit price or allowance. Adobe’s free allocation counts Document Transactions, so dividing that allowance by physical pages would be wrong. Record which steps you actually call: parsing, extracting fields, splitting, embeddings, model inference and storage may sit on different bills. This site’s AI API cost calculator can estimate downstream model-token charges; it is not a document-parser meter.
Price a mixed-length workload
Here is an illustrative workload, not observed traffic or a vendor quote: 100 one-page invoices and 100 eleven-page invoices. That is 100 × 1 + 100 × 11 = 1,200 physical pages in 200 files. With Google’s invoice parser, the first group consumes 100 × ceil(1/10) = 100 billable blocks. The second consumes 100 × ceil(11/10) = 200 blocks. Total: 300 blocks × $0.10 = $30 before other charges. Dividing 1,200 pages by ten first gives 120 blocks and $12, but that undercounts the published per-document rule by $18. Google’s pricing examples establish both the block size and the $0.10 first rate.
The same 1,200 physical pages at Google’s listed $30-per-1,000-page Form Parser first tier would be $36 before any applicable allocation or other service cost. That arithmetic compares tariffs, not invoice-parser and Form-Parser quality or suitability. A Google invoice-parser document above ten pages requires batch processing under the page’s stated synchronous limit; Google’s pricing page says batch can process documents up to 200 pages for that specialized route. Check the processor and method you would actually deploy.
Page selection can change a different bill when the API supports it. Azure’s pricing FAQ says all submitted pages are analyzed unless the request specifies a range with pages. Suppose, illustratively, that only pages 1–2 of each eleven-page file hold the fields you need. Selecting those two pages would make the workload 100 × 1 + 100 × 2 = 300 analyzed pages if that selection is valid for your document and model. It is not a claimed Azure dollar saving: this article did not establish a paid Azure rate. Keep the original files and verify that omitted pages never carry required fields or context before routing a production workload this way.
Decide when a second pass pays off
A cheap first pass followed by a stronger pass for exceptions has a clear API arithmetic, but it is useful only if you measure exceptions and the outputs satisfy the same task. Let P be pages, rA the first-pass price per page, rB the second-pass price, and f the fraction sent to the second pass. Then API cost = P × rA + f × P × rB. An all-premium route would be P × rB, so the two-pass API spend is lower only when f < 1 - rA/rB, assuming the two published rates cover equivalent billable pages and no minimums or extra features apply.
For a labelled illustration, use Google’s OCR first-tier $0.0015/page and Layout Parser $0.01/page, with 1,200 pages and a measured-in-your-own-workflow exception fraction that you assume to be 20% for planning. The API-only model gives 1,200 × $0.0015 + 240 × $0.01 = $4.20, versus 1,200 × $0.01 = $12 for all-layout. The mathematical threshold is 85%. These are tariff calculations, not a claim that OCR identifies layout exceptions accurately, that the processors can be substituted, or that this workflow saves $7.80 in practice. Google pricing supplies the rates; only a representative evaluation can supply f and acceptable output quality.
Add review work explicitly: review cost = reviewer minutes / 60 × assumed hourly rate. For example, ten illustrative minutes at an assumed $30/hour adds $5, not a published wage fact. Also record retry charges under the chosen API’s terms, downstream tokens, storage and any separate extraction call. The AI API price comparison helps with token tariffs; the web search cost guide shows how separate retrieval calls can accumulate, and the speech-to-text cost guide uses the same principle of counting native billing units before multiplying by a rate.
Check limits before committing
Run a representative sample through the exact route and record what comes back: missing tables, split rows, handwritten text, field completeness, latency and exception rate. This article reports no measured accuracy for any provider. A vendor’s pricing page cannot prove how well it handles your files. Keep the selection rule and acceptance criteria the same across candidates.
Check document size, page cap, sync versus batch rules, concurrency and retention for the chosen plan. For example, LlamaParse limitations lists a 512 MB Parse file limit and a default 10,000-page limit after page selection. Google’s pricing page states the invoice-parser synchronous ten-page boundary and 200-page batch maximum. Those limits belong to named routes, not every product from either provider. Also check whether a free allocation expires, requires a new account, or applies to the paid route you intend to use. Then request a region-specific or contract-specific quote where the public figure is missing.
Sources
- Amazon Textract pricing — OCR, feature combinations, Oregon examples and free tier.
- Google Document AI pricing — OCR, layout, form and invoice-parser rates, block examples and limits.
- Google extraction overview — output distinctions.
- Azure Document Intelligence pricing — public free allocation, page-selection FAQ and unreadable paid figures.
- Mistral pricing — OCR 4.1 page rate.
- LlamaParse pricing, billing and limits — tier credits, purchase rate and file limits.
- Reducto pricing — Standard Parse and Extract rates.
- Adobe Acrobat Services pricing and PDF Extract overview — transaction allocation and product output.
All pages read on November 7, 2026.
Free toolWhat will 10,000 chats a month cost?Estimate my bill
