Research-backed calculator guide

Table Extraction API Cost per Page vs Cost per Usable Table

Compare table extraction API cost per page with the production cost per usable table, using provider units, correction yield, retries and a worked calculation.

What “cost per page” actually measures

A page is a billing unit, not an outcome. One page may contain no table, one small table, or several dense tables. A scanned page may also require more review than a clean digital PDF. Before comparing table OCR pricing, define what your team calls a successful result. A usable table might need the expected row and column count, correct headers, numeric values that reconcile to a source total, and a machine-readable output accepted by the next system.

A provider’s page rate generally describes an API operation. It may include OCR, layout analysis or table structure, but it does not necessarily include storage, document download, queueing, post-processing, quality checks, human correction, retries or the engineering time required to maintain the integration. Mistral’s OCR documentation illustrates why output details matter: its processor can return structured content, table formatting, page metadata and confidence scores, and its response exposes pages and table information. ^[5] Those features can reduce downstream work, but whether they do so is a workload question to measure rather than a guaranteed saving.

The practical distinction is:

• Processed-page cost answers: “What did the extraction provider charge for requests covering these pages?”
• Detected-table cost answers: “What did the extraction run cost divided by tables returned?”
• Usable-table cost answers: “What did the full workflow cost divided by tables accepted without unplanned rework?”

Use the first for reconciling a provider invoice. Use the third for a build-versus-buy decision, a customer price, or a margin model.

A provider-rate reality check

Published pricing pages expose different units. AWS provides a direct table feature example. Its page lists Analyze Document Tables at $0.015 per page for the first one million pages in the cited US West (Oregon) example and $0.010 per page after that threshold. It also shows that combining Tables with Forms or Queries changes the price. ^[1] If your request asks for more than tables, do not copy the table-only rate into the model.

Google Cloud’s Document AI page uses processor pricing. Its examples show Form Parser at $30 per 1,000 pages for the first million pages and $20 per 1,000 pages above that tier. The page also explains that a “page” depends on input format, such as one PDF page, one image, one TIFF image, or a page-equivalent for some office formats. ^[3] A table workflow using another processor or add-on needs its own line item; the Form Parser number is not a promise of a generic table price.

Azure Document Intelligence shows a different menu of pay-as-you-go rows. The page lists Read at $1.50 per 1,000 pages for the first million pages and $0.60 thereafter, all prebuilt models at $10 per 1,000 pages, custom extraction at $30 per 1,000 pages, plus separate query and add-on rows. It states that pages are analyzed unless a page range is specified. ^[4] Treat those rows as dated estimates of the selected service configuration, not as a single “Azure table price.” Microsoft also notes that displayed prices are estimates and may vary by agreement, purchase date, region and currency. ^[4]

These published examples are useful for building a comparable worksheet, but they do not establish that one API will be cheaper for your documents. A fair comparison fixes the same page sample, table-bearing rate, required fields, acceptance test, retry policy and labor assumptions. It also records whether each provider returns tables in JSON, Markdown, HTML, CSV or a layout-specific structure. Mistral, for example, documents Markdown and HTML table formatting options and confidence-score granularity, which can affect validation design even when a published price is not the only decision variable. ^[5]

The cost-per-usable-table formula

Use these variables for a first-pass model:

• P = pages submitted.
• r = API price per processed page, in USD/page.
• q = table-bearing pages as a share of all pages.
• t = average tables per table-bearing page.
• a = acceptance rate after automated validation, expressed as a decimal.
• x = share of pages retried or sent to a fallback path.
• r2 = average cost per retried or fallback page.
• m = share of rejected tables requiring manual correction.
• s = correction seconds per rejected table.
• w = loaded reviewer cost per minute.
• E = allocated engineering and operations cost for the period.

Expected raw tables are P × q × t. Expected usable tables are:

U = P × q × t × a

Base API cost is P × r. Retry or fallback cost is P × x × r2; if the fallback is free, set r2 to zero, but record that assumption. Correction cost is:

C = P × q × t × (1 − a) × m × (s ÷ 60) × w

The modeled total is:

T = (P × r) + (P × x × r2) + C + E

Finally:

cost per usable table = T ÷ U

Do not multiply correction cost by every rejected table if your process discards some outputs without human review. Conversely, do not set correction to zero merely because a vendor returns confidence scores. Confidence can support triage; it is not the same thing as your acceptance test.

Worked cost example

The following is an original planning example, not a vendor quote. It uses the published AWS table rate described above only as the API-rate input. The table makes the assumptions and outputs inspectable.

The API line alone is $750. The full modeled workflow is 2.51 times that amount because this workload has correction and engineering costs. That multiplier is not a quality score and should not be transferred to another provider without remeasuring the inputs.

Notice also that the page rate and the usable-table rate answer different questions. The provider charges by page, while the business receives accepted tables. If the sample has fewer table-bearing pages, the cost per usable table rises even when the API price does not change. If a page has more tables and the acceptance rate stays high, the denominator grows and cost per usable table falls. This is why “cheapest per page” is not a sufficient selection rule.

How to measure the missing inputs

Start with a labeled sample, not a guess. Select documents that represent the production mix: native PDFs, scans, rotated pages, multi-page tables, merged cells, subtotals, handwritten marks and multilingual content if those occur in the queue. Record pages, table-bearing pages, table count, output format and the specific fields that must reconcile.

Measure acceptance at the table level. A table can have correct text but incorrect row boundaries. It can preserve rows while shifting a value into the wrong column. It can produce a valid JSON object that fails a downstream total check. Create deterministic checks for schema, required headers, row counts, numeric parsing, totals and duplicate rows. Then calculate a from the sample.

Measure correction seconds from a reviewer’s real workflow. Include opening the evidence, locating the error, editing the output, recording a reason and submitting the correction. If different table classes have different durations, calculate a weighted average or model them separately. The $0.50/minute example is an estimate; use a loaded rate that includes the portion of wages, benefits, vendor fees or internal labor you intend to attribute.

Measure retries separately from ordinary volume. A timeout, malformed output, provider error and a business-rule rejection are different events. A retry caused by transport failure may incur another API charge. A business-rule rejection may route to a fallback model or to a person. Keep attempt IDs and provider usage records so the AI API invoice reconciliation calculator can compare modeled requests with billed units.

If table extraction feeds a retrieval system, add downstream costs separately. A normalized table may require chunking, embedding, vector storage and query-time retrieval. The RAG, embedding and vector database cost calculator is the appropriate next worksheet for those lines. If you are comparing multiple model or provider configurations, use the Model pricing comparison route to keep units and assumptions visible rather than mixing page rates with token rates.

Sensitivity: what changes the answer fastest?

Acceptance rate and table density usually dominate the denominator. In the worked sample, reducing acceptance from 92% to 80% lowers usable tables from 33,120 to 28,800 while leaving the $750 base API charge unchanged. If correction remains 30 seconds per rejected table, correction cost also increases. Raising table-bearing pages from 60% to 80% increases expected raw and usable tables, but only if the new pages have similar quality and table count.

Engineering allocation matters most at low volume. If the same $400 monthly allocation is spread over 5,000 pages instead of 50,000, it adds eight cents per page before quality labor. At high volume, a fixed allocation may become small per page, but monitoring, schema changes and exception handling can become variable costs. Keep fixed and variable lines separate so the model can show both a pilot and a steady-state case.

Volume tiers can change the API rate. AWS, Google and Azure each publish tier or processor distinctions, but the thresholds and units differ. ^{[1] [3] [4]} Run a separate case for each threshold rather than applying the lower marginal rate to all pages. Include free tiers only when the workload is eligible and the allowance is available in the billing account; do not treat trial credits as a permanent unit cost.

Exclusions and limitations

This article’s worked number excludes storage, data transfer, queue infrastructure, taxes, minimum commitments, support plans, observability, security review, human adjudication for uncorrectable tables and the cost of collecting a labeled test set. Add those lines when they are material. It also assumes one billing period, a stable document mix and a single average retry rate.

Vendor pages change. The rates and examples above were checked on 2026-11-02 and are quoted with their stated units or assumptions. A page’s current price, region, contract, processor version or free-tier eligibility may differ when you run the calculation. Use the provider’s current pricing page and your invoice as the authority for a purchase decision. The search benchmark also cannot prove that every returned listing was organic because the search interface did not reliably expose ad labels.

A cost model does not measure extraction quality by itself. Validate the same sample with the same acceptance rules, publish the confusion cases, and report both page cost and usable-table cost. If the table is used for regulated reporting or a financial decision, keep an evidence link and a human escalation path even when the arithmetic looks favorable.

FAQ

Is table extraction API cost per page the same as cost per table?

No. Page billing counts submitted or analyzed pages. Cost per table divides the workflow cost by detected or accepted tables. A page with no table still consumes a page unit, while a dense page may produce several tables. Use cost per usable table when the output, not the document page, is the unit that creates value.

How do I calculate PDF table extraction cost for a monthly workload?

Multiply the pages that will actually be analyzed by the selected provider rate, then add retry or fallback requests, correction labor and any fixed allocation. Divide the total by expected usable tables. Keep the provider’s billing unit and your table-yield assumptions in separate fields so a change in document mix does not silently change the quoted API rate.

What should I include in table parsing API pricing comparisons?

Compare the same document sample, table-bearing share, tables per page, acceptance test, output format, retry policy and labor rate. Record whether the provider charges for add-ons, forms, queries, layout, storage or batch processing. Google and Azure publish processor or feature-specific rows rather than one universal table price, so use the exact configuration and unit shown on the current pricing page. ^{[3] [4]}

Can confidence scores be used as the acceptance rate?

Not automatically. A confidence score is a signal for routing or review. Acceptance should be measured against your required schema, row and column structure, numeric checks and downstream reconciliation. Some high-confidence tables can still have structural errors, while a lower-confidence result may be correct after review. Estimate acceptance from a labeled sample and document the threshold.

Should retries be included when a provider charges per page?

Yes. If a request is retried and the provider bills the second attempt, it is additional page volume. Keep transport retries, provider errors, business-rule rejections and fallback calls as separate counts. If your logs show that some failed requests are not billed, use that observed rule and cite the provider’s billing terms; do not assume all failures are free.

When is cost per usable table better than cost per page?

Use it when comparing providers, pricing a document-processing feature, setting a margin target or deciding whether manual correction is affordable. Cost per page remains useful for invoice reconciliation and capacity planning. Reporting both prevents the workflow from looking inexpensive simply because the denominator stops at pages processed.