Research-backed calculator guide

OCR API Cost per Page: Compare the Meter Before You Budget

Learn how to normalize OCR API cost per page across image, PDF-page, document, request, credit, and batch meters, with formulas and a worked budget example.

What “cost per page” should mean

A page is a physical or logical unit of document content, but vendors do not all define it in the same way. A raster image may be one page. A PDF page may be one page. A Word file, spreadsheet, slide deck, or HTML file may be converted into pages using a provider-specific rule. Google Document AI’s pricing notes, for example, define an image as one page, a PDF page as one page, a TIFF image as one page, an Excel tab as one page, and a PowerPoint slide as one page. It also uses a character-based rule for some text formats.^[2]

That definition matters before you compare an OCR API pricing per 1,000 pages. Ask these questions:

1. Is the meter based on pages, images, documents, requests, credits, or subscription capacity?
2. Does a multi-page file count once or once per page?
3. Is OCR alone included, or are tables, forms, layout, handwriting, fields, or queries separate features?
4. Is the listed rate marginal by tier, blended across the month, or available only through a committed contract?
5. Is there a free allowance, a prepaid minimum, or a monthly platform fee?
6. Does asynchronous batch processing change price or only change delivery time?

AWS Textract illustrates why the feature question is important. Its Detect Document Text API uses OCR for text and handwriting, while Analyze Document adds Forms, Tables, Queries, and Signatures. The pricing page lists separate per-page examples for these feature combinations, and the OCR is included in Analyze Document output rather than being a second copy of the base text charge.^[3] The exact region and tier therefore belong in the estimate.

A provider-neutral OCR cost formula

Use this sequence for a document OCR cost calculator or spreadsheet.

Step 1: Convert files to pages.

billable pages = number of files × average pages per file

For mixed files, calculate each class separately. For example, 1,000 PDFs at four pages each plus 500 single-page images equals 4,500 pages. Do not use the file count as a proxy unless every file is known to be one page.

Step 2: Separate required services.

Keep plain OCR, layout, tables, forms, named-field extraction, classification, translation, storage, and review as separate rows. A vendor may bundle some of them, but modeling them separately makes the estimate auditable and lets you replace a rate when the provider changes its packaging.

Step 3: Apply the billing curve.

For a flat rate, multiply pages by rate. For marginal tiers, split the pages at each threshold:

processing cost = (pages in tier 1 × tier 1 rate) + (pages in tier 2 × tier 2 rate) + ...

Subtract a free allowance before applying paid tiers only when the vendor documents that the allowance is shared with the exact OCR feature you are using. Google Cloud Vision states that the first 1,000 units per month are free for the listed feature and that the final 1,000-unit block is prorated.^[1] That is not automatically transferable to another Google service or another vendor.

Step 4: Add workflow costs.

monthly OCR budget = processing + optional features + retries + review + storage/operations

Keep one-time integration work separate from recurring usage. If you need a blended unit cost, divide the recurring total by the same billable page count used by the API estimate.

Worked cost example

Assume a US English-language document workflow checked on 2026-11-02. The workload contains 2,000 PDF files, each averaging five pages. The application needs plain OCR only. Use an illustrative rate of $4 per 1,000 pages, matching the OCR 4.1 list rate shown by Mistral’s official pricing page on that date.^[4] This is a modeling assumption for the example, not a promise that every request will receive that rate.

The API-only estimate is $40, or $0.004 per page. Once the assumed review work is included, the recurring subtotal is $160, or $0.016 per page. The difference is the reason to report both API cost per page and fully loaded workflow cost per page. This example excludes storage, network egress, retries, failed-request policy, taxes, implementation, and structured extraction.

Now change only the file shape. If the same 10,000 pages arrive as 10,000 one-page images, the page total stays 10,000. If a provider bills per document instead of per page, however, the file count and pages-per-file distribution may change the result. A five-page PDF billed at $0.02 per document is $0.004 per page for that file, while a one-page document at the same document rate is $0.02 per page. Document size is therefore an economic variable, not just a processing detail.

How to compare published OCR rates fairly

Start with the simplest comparable row: plain OCR of the same page type. Cloud Vision lists Text Detection and Document Text Detection at $1.50 per 1,000 units for the 1,000-to-5,000,000 range on the checked pricing page, while noting that a multi-page PDF is charged page by page as individual images.^[1] That is a useful page-normalized baseline only if you are comparing the same feature and not adding another feature to the request.

Google Document AI uses a different product shape. Its Enterprise Document OCR Processor is listed at $1.50 per 1,000 units in the corresponding range, while some specialized processors and add-ons have separate charges. Its examples also show that certain specialized processor counts can represent up to 10 pages per count, so “per count” is not always “per page.”^[2] Treat the vendor’s definition as part of the rate card.

AWS provides another comparison point. The Textract pricing page gives a US West (Oregon) example of $0.0015 per page for the first million pages for Detect Document Text, while Forms, Tables, Queries, and other Analyze features use different examples and tiers.^[3] Do not combine a plain OCR quote with a structured extraction requirement and call the result a like-for-like comparison.

Mistral’s official table uses a direct page denominator for OCR models. The page lists OCR 4.1 at $4 per 1,000 pages and OCR 3 at $2 per 1,000 pages, with cached-input columns shown in the OCR model table.^[4] If your workload is eligible for another service tier or commercial arrangement, use that rate instead of the public list rate and label it as a quote or contract assumption.

A subscription plan can be harder to normalize. OCR.Space’s public API page, which appeared in the checked results, presents free and paid monthly plans with request quotas and PDF page limits rather than a simple page rate. Microsoft Purview’s OCR Cost Estimator similarly focuses on scanning selected workloads and projecting costs before deployment; it says the first estimation can take around 24 hours and documents scope limitations, such as different treatment of embedded images across workloads.^[5] These are useful reminders that “OCR cost per document” may include a capacity or platform model rather than a pure variable page charge.

What to include beyond the API line item

Optional extraction. Named fields, tables, forms, queries, and classification can be priced separately or bundled into a higher feature rate. Model them as explicit rows.

Retries and duplicates. A timeout can lead to a retry that is billed even if your application eventually discards the first response. Use an observed retry rate from logs when you have it. Until then, show a scenario rather than presenting a made-up certainty.

Human review. Low API spend can still produce a high unit cost if difficult scans require correction. Multiply expected review hours by a documented internal or vendor labor assumption.

File-shape distribution. Averages hide skew. Track p50 and p95 pages per file if large PDFs or long archives are common. A provider that bills per document can produce a different result from one that bills per page.

Batch and latency. A batch discount is valuable only if its completion window fits the workflow. Keep interactive and asynchronous rates in separate scenarios.

Minimums and commitments. A monthly minimum, prepaid credit, or reserved capacity can turn low-volume effective cost into a higher figure than the nominal page rate. Allocate that fixed amount over expected pages and show the volume assumption.

For a broader AI spend view, compare this page-level model with the model pricing comparison, reconcile estimates against actual usage in the AI API invoice reconciliation calculator, and include retrieval or embedding work with the RAG, embedding and vector database cost calculator. These are contextual links, not substitutes for a provider’s current rate card.

FAQ

How much does OCR API cost per page?

There is no single market-wide rate. On the official pages checked on 2026-11-02, examples ranged from low single-digit dollars per 1,000 pages for some plain OCR meters to much higher per-page rates for structured extraction features. Your exact cost depends on provider, region, tier, document type, and features. Use the provider’s current rate card and calculate pages first.

Is OCR priced per page or per document?

Both models exist. Cloud Vision says each page in a multi-page PDF is treated as an individual image for billing.^[1] Other services publish document, request, credit, subscription, or mixed feature meters. Ask what happens to a five-page PDF and a one-page image before accepting a “per document” comparison.

How do I calculate OCR API pricing per 1,000 pages?

Divide the published 1,000-page rate by 1,000 to get a page rate, then multiply by your billable pages. For a $4 per 1,000-page list rate, the calculation is $4 ÷ 1,000 = $0.004 per page; 10,000 pages cost $40 before optional charges. Apply free allowances and marginal tiers only as documented by that provider.

Does PDF length change the OCR cost?

It can. With page-metered OCR, a longer PDF creates more billable pages. With document-metered pricing, the number of pages may change processing limits or effective unit economics even when the nominal document fee is fixed. For each file class, record file count, pages per file, and the billing unit.

Does OCR include tables and field extraction?

Not necessarily. Plain OCR may return text while tables, forms, layout, queries, handwriting, or named fields use a separate feature or processor. AWS Textract’s pricing page explicitly separates Detect Document Text from Analyze Document features such as Forms, Tables, Queries, and Signatures.^[3] Treat “OCR included” as a feature statement, not proof that every downstream extraction is free.

Should I include human review in cost per page?

Report two figures: API-only cost per page and fully loaded workflow cost per page. API-only is useful for vendor comparison. Fully loaded cost adds review, retries, storage, and operations, and is more useful for a product or finance budget. Label labor rates and hours as assumptions unless they come from your own measured workflow.

Are free OCR pages safe to subtract from every estimate?

No. Confirm the free allowance’s scope, period, feature, account eligibility, and whether it is shared with other products. Some plans use request quotas or monthly subscriptions rather than free pages. The official page should control the calculation.

How often should I refresh an OCR estimate?

Refresh whenever the provider changes its rate card, model, processing mode, or tier definition. Also refresh when your file mix changes, because page distribution, extraction requirements, and review rates can move the fully loaded cost even when the API list price does not.

Assumptions and limitations

This guide uses USD list-price observations available on 2026-11-02 and a benchmark locale of English, United States. Rates can change, and regional taxes, currency conversion, negotiated contracts, credits, committed-use discounts, and account-specific eligibility can change the payable amount. Official pricing pages are factual references, not quotes for your account.

The worked example assumes five pages per PDF, a flat illustrative OCR rate, four review hours, and $30 per review hour. It does not estimate search volume, traffic, ranking, revenue, or savings. It also does not claim that one provider is universally cheaper or more accurate. Validate page definitions, feature scope, latency, retention, and quality on representative documents before committing to a budget.