How many tokens is a PDF page? Claude vs Gemini

Name: PDF token cost per page: Claude vs Gemini, and PDF vs pasted text
Creator: TokenSave
Published: 2026-10-18
License: https://creativecommons.org/licenses/by/4.0/

One page of text costs 2,347 tokens as a PDF on Claude 5.5 but 728 when the same text is pasted (+222%). Gemini charges a flat 560 tokens per PDF page whatever is on it, so the same page costs 76% less there.

Measured 2026-10-18 · text, table and scanned pages · 1 to 10 pages · License CC BY 4.0 · Related: image token cost

DOI 10.5281/zenodo.23446565 · Also on Hugging Face · Kaggle · GitHub

⬇ CSV ⬇ JSON How to cite

Input tokens for one page as a PDF or as pasted text on Claude, Gemini and GPT

Key findings

Claude reads every PDF page twice: as text and as an image. A text page costs 2,347 tokens on Claude 5.5. A scanned page with no text layer costs 1,600, which is the image part alone; the rest is the extracted text.
Pasting text is much cheaper on Claude. The same page pasted as text costs 728 tokens on Claude 5.5, 69% less than the PDF. If you only need the words, copy them out.
Gemini charges a flat rate per page. Text, table and scanned pages all cost 561 tokens, and 10 pages cost 5,601, so about 560 per page.
On Gemini a PDF can beat pasting. Our table page costs 561 tokens as a PDF but 777 pasted as CSV (28% less as a PDF). For dense pages Gemini's flat rate is cheaper than the text itself.
Cost grows in a straight line with pages. 10 text pages cost 23,236 tokens on Claude 5.5 (about 2,324 per page) and 20,979 on Claude Haiku 4.5.

All inputs

Input tokens. GPT is shown for pasted text only (see method).

Input	Claude Opus/Sonnet 5.5	Claude Haiku 4.5	Gemini 3.x	GPT o200k
PDF, text, 1 page	2,347	2,115	561	–
PDF, text, 5 pages	11,631	10,499	2,801	–
PDF, text, 10 pages	23,236	20,979	5,601	–
PDF, table, 1 page	2,481	2,278	561	–
PDF, scanned image, 1 page	1,600	1,592	561	–
Same page pasted as text	728	503	455	461
Same table pasted as CSV text	913	720	777	629

Method

A4 PDFs made with ReportLab: one page of English prose (about 420 words), the same page repeated to 5 and 10 pages, a page with a 35-row table, and a scanned version of the text page (rendered at 150 dpi and placed as an image with no text layer).
Claude: each PDF sent through Claude Code (claude -p, stream-json input) as a document block with a one-character message; input tokens minus the same run without the PDF. Opus 5.5 and Sonnet 5.5 gave identical counts.
Gemini: per-file token count shown by Google AI Studio for Gemini 3.8 Flash after uploading each PDF. Pasted text was measured with the prompt counter minus its one start token.
GPT: OpenAI bills PDF input as extracted text plus an image of each page, which we could not reproduce without the API, so GPT appears only for pasted text (tiktoken, o200k_base).

Limitations

One page design. Pages with more text, small fonts or many images will cost more on Claude; Gemini's per-page rate should stay flat.
Default settings; Gemini's media resolution option changes its per-page count.
Input tokens only; the price per token differs by model.

How to cite

Free to use, quote and republish with attribution under CC BY 4.0. Please link to this page.

TokenSave (2026). PDF token cost per page: Claude vs Gemini, and PDF vs pasted text [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23446565

Embed the chart

<a href="https://tokensave.app/data/pdf-token-cost"><img src="https://tokensave.app/pdf-token-cost-chart.png" alt="Input tokens for one PDF page on Claude and Gemini, and the same page pasted as text (TokenSave)" width="1200" loading="lazy"></a>
<p>Source: <a href="https://tokensave.app/data/pdf-token-cost">TokenSave, PDF token cost per page</a> (CC BY 4.0)</p>

Questions or corrections: contact@jonhisking.com