← Back to articles

Document understanding · API billing

How much does it cost to analyze 1,000 PDF pages with an LLM?

PDF page count alone is not a reliable bill. OpenAI, Claude and Gemini turn document content into billable model input differently. Start with actual API usage when available; otherwise make the token assumptions visible.

Published Nov 16, 2026Official sources checked Nov 16, 2026

Provider rules

One PDF can produce text and image input

OpenAI

Vision-capable models process PDF files as extracted text and page images. In the Responses API, input_file.detail can be low, high or auto, changing page-image processing but not excluding extracted text. Other API endpoints have different controls.

Official file-input guide

Anthropic Claude

Claude extracts text and renders each PDF page as an image. Its guide gives roughly 1,500–3,000 text tokens per page for typical pages, plus image tokens. It says standard API pricing applies, without a separate PDF fee. That text estimate is not an all-in per-page limit.

Official PDF support and cost guide

Google Gemini

Google's document guide describes a 258-token page equivalent, with model-specific resolution behavior. For Gemini 3, native extracted PDF text is not charged, while processed page-image tokens appear under IMAGE in usage metadata. Do not apply earlier-model document accounting to these responses.

Official document-processing guide

A per-page OCR service can use an entirely different pricing unit. See page-priced OCR and image-input token rules before comparing unlike services.

Most accurate path

Estimate from the API response's billable usage

  1. Process a representative sample of your real PDFs, including a scan and a text-rich document if both occur in production.
  2. Read the provider's total billable input and output token counts and any billed cache, batch or media categories. Do not add image tokens again when they are already included in the billed input total.
  3. Match the exact model and API route to its current input/output rates in the model catalog. Confirm tier and any special billing rules with the provider.
  4. Use the token calculator with those totals and your expected number of requests; account separately for retries, repeated PDF analysis, and charges outside token billing.

For example, suppose an actual batch of 1,000 PDF pages reports 2.4 million billable input tokens and 80,000 billable output tokens. At hypothetical rates of $1 per million input and $4 per million output, the calculation is 2.4 × $1 + 0.08 × $4 = $2.72. These rates are arithmetic examples, not prices quoted for OpenAI, Claude or Gemini. If the same workload instead bills 4 million input tokens, the total at these assumed rates is $4.32.

Before a pilot

Budget without measured tokens

Count documents, pages per document, questions per document and expected output length. For illustration: 100 PDFs with 10 pages each is 1,000 pages. If your assumed total billable input is 1,800 tokens per page plus 500 prompt tokens per PDF, and output averages 700 tokens per PDF, then input is 1,850,000 tokens and output is 70,000 tokens. At the same hypothetical $1/$4 per-million rates, that would be $2.13. This is a budgeting scenario, not an estimate of any provider's PDF tokenizer.

Actual costs can differ with scanned versus searchable PDFs, page layout, image detail/resolution, provider/model generation, instructions, caching, batch eligibility and failed requests. Gemini 3's native-text rule makes cross-provider page-to-token assumptions especially unsuitable. Validate with measured usage before estimating a large workload.

This site's calculator prices supplied token counts; it does not parse or upload PDFs or calculate PDF image tokens for you. For the surrounding costs, see hidden LLM API costs.