How Many Tokens Is a PDF in Claude? 195 PDFs, Uploaded vs Pasted as Text

Anthropic says how a PDF is billed: the text of each page, plus an image of each page. It does not say what that comes to on the PDFs people have. So we took 195 PDFs from the public web and counted each one twice with Anthropic's own token counter, once uploaded as a PDF and once as the text FileConcat extracts from it.

About 2,400 tokens for a typical page uploaded as a PDF, and about 900 for the same page pasted as text. The difference is the picture of the page that Claude reads alongside its text.

That price per page is fixed by the image, not by how much the page says. A page with one line of text still pays for its whole picture, so short pages cost the most extra. A dense page of small print is the one case where the text alone can outweigh the image: two of the 165 PDFs cost more as text.

Across whole files the saving holds up. Pasted text costs a third of the upload or less for close to half the PDFs, and half or less for nearly nine in ten.

When the page's picture carries something its words do not. The image is what lets Claude read a chart, a photo, a diagram or a scanned page. Text keeps the words and leaves all of that behind, and two in three of these PDFs hold at least one embedded picture, even if many of them are only a logo.

• Paste the text when the question is about the words. Contracts, reports, papers, manuals and forms are read from their text, and about three fifths of the upload's tokens pay for pictures of pages the answer does not need.
• Upload the PDF when the answer is in a figure. A chart's values or a diagram's labels are often not in the text layer at all. Check your text before you trust it for that.
• A scanned PDF has no text to paste. FileConcat names the pages that are images and offers to read them on your machine. In the claude.ai apps, a PDF past 100 pages is read as text only anyway.

The saving repeats in a long conversation, because the document stays in it and is read again with every message. Our measurement found two PDFs where the text costs more, and both are single pages packed with small print.

The figures above are a distribution. Your PDF is one point in it, and the result screen shows its token count before you paste anything.

Drop a PDF here and FileConcat extracts its text in your browser, tells you which pages are scanned, and counts what the text costs. Your files are not uploaded to us. What you then paste into an assistant is your own choice, but preparing the file uploads nothing.

Anthropic's own PDF documentation gives the rule and a range, not a measurement. It says "Each page typically uses 1,500-3,000 tokens depending on content density" for text, and "Because each page is converted into an image, the same image-based cost calculations are applied". It then points the reader at token counting to find out for their own PDFs. That is what this page does, for 195 of them.

The text range does not describe these PDFs. The median page here extracts to 856 tokens of text, and the median page costs 2,390 as a PDF all in. Real web PDFs are short and light: the median file is 3 pages and about a third are a single page.

The one study we found that logs Claude's PDF token counts is Ak et al. (arXiv 2609.29933). On Claude Sonnet 4.5 it saw a mean of about 2,050 tokens per page on 50 to 100 page documents, falling to about 1,540 above 100 pages, which it attributes to the API downsampling page images to fit a 200K window. It never counts the same documents as plain text, so it has no ratio.

It also says that without citations the API "falls back to basic text extraction". Our counts contradict that for the direct API. Every count here was made without citations, and every page still cost about 1,600 tokens more than its text. That gap matches what a page with no text at all costs, as the scanned group below shows. We counted only the direct API, so the fallback may belong to another platform the paper does not name.

Nobody had counted the same real PDFs both ways with the vendor's own counter. That is this page.

The PDFs are a seeded draw (seed 20261016) from the Common Crawl crawl CC-MAIN-2026-34: documents people published on the open web. The crawl's index picks records by detected MIME type, at most two files per registered domain so one site cannot fill the sample, byte-identical files once. Records the crawler cut short are skipped, which leaves out the largest files. The builder is packages/cli/scripts/build-cc-corpus.py, and the same Common Crawl draw is the one the extractor comparison weights its scores by.

Your PDFs are not on the web, so this is a stand-in for them, not a sample of them. Every file was kept, whatever it measured. FileConcat's reader sorted the 195 into four groups before anything was counted:

The headline numbers are the text-layer group, because only there is the text a fair substitute for the page's words.

1. 1

   Extract the text the way the browser does

   run-corpus.sh runs every PDF through the web app's own reader chain, one process per file, and keeps the text it produces. This is what FileConcat puts in a bundle, page markers included.

2. 2

   Count the PDF as Claude receives it

   The file goes to the token counting endpoint as a base64 PDF document block in one user message, exactly as an upload through the API would. The endpoint is free and counts under the tokenizer of the model named.

3. 3

   Count the text the same way

   The extracted text goes to the same endpoint as one plain text block, in the same one-message request.

4. 4

   Subtract the envelope

   A one-message request costs 9 tokens before any content. That is counted once and subtracted from both sides, so each figure is the content alone.

5. 5

   Count pages and pictures with pdf.js

   The web app's PDF library gives each file's page count and the pages that paint an embedded raster image. Drawings made of vector paths are not counted as pictures, so that figure is a floor.

Pooled, the 165 PDFs count 4,432,664 tokens uploaded and 1,674,624 as text: 2.65 times as much, or 62.2% saved by pasting the text. Per file the median ratio is 2.81x, the 10th percentile 1.78x and the 90th 5.64x. The pooled figure sits below the median because pooling weights each file by its size, and the files with the most text per page are where the gap is narrowest.

The 100-page cut matters because claude.ai reads a PDF past 100 pages as text only, and because the API caps a request at 100 pages when the model's window is under a million tokens. Three files is too few to say anything about long PDFs beyond this: their median page counted the same 2,390 tokens as a short file's, so on this model we saw none of the downsampling the arXiv study logged.

The PDF's extra cost per page has a median of 1,568 tokens over the text. On the scanned PDFs, where most pages have no text layer, the median page costs 1,586 as a PDF. The two numbers are close because both are mostly the same thing: the image of one page.

Where the text costs more. Two files: c9a5be22.pdf, one dense newspaper page at 10,238 tokens as text against 5,247 as a PDF, and 94ba3f01.pdf at 0.99x. Why the newspaper page's PDF counts about half its extracted text is not verified. One reading is that Claude's own text extraction of that page is shorter than ours; we did not see the text the endpoint extracted, only its count.

What text leaves behind. 112 of the 165 hold at least one embedded raster image, on 551 of their 1,807 pages. Many are logos and decorations, and vector charts are not counted at all, so this is how often a page has a picture, not how often the picture matters.

Scanned, 24 files. The PDFs whose pages are images cost 745,336 tokens uploaded, 1,586 for the median page. FileConcat's reader extracted text from 5 of them, the pages that have a text layer, and nothing from the other 19. Pasting their text is not a saving, it is a loss of the document. FileConcat names those pages and offers recognition on your machine instead, which this measurement did not run.

Undecodable, 5 files. Part of their text layer uses a font with no usable character map, so some of the extracted text is not words. They cost 174,582 tokens uploaded and 84,366 as text, 2.07x, and are left out of the headline because the text side is not a fair substitute.

Rejected, 1 file. dba5166a.pdf is 4 bytes. The endpoint answers "The PDF specified was not valid". It is counted in the 195 and nowhere else.

• This is the Claude API, not the claude.ai apps. The apps do not publish what a PDF costs in tokens. Their help page says a PDF of 101 to 1,000 pages is read as text only, which is the text side of this page done for you. Whether a shorter upload in the apps costs what the API counts is not verified.
• One model, one tokenizer. claude-opus-5-5 shares the tokenizer introduced with Claude Opus 4.7. Anthropic says that tokenizer counts roughly 30% more than earlier models. Both sides move together, so the ratio should hold on older models, but we did not count them.
• Web PDFs are short. The median file is 3 pages. The largest files were never in the sample, because the crawl does not keep files it cut short.
• Cost, not answers. This counts what reaches the model. It does not say whether the answer is as good. Ak et al. found raw PDF the strongest input for Sonnet 4.5 on a long-document benchmark, and extracted text the cheapest. Losing the picture is a real price for a question about a figure.
• The picture count is a floor. pdf.js reports pages that paint a raster image. A chart drawn as vector paths is not counted, and a logo counts the same as a photo.
• The bundle wrapper is not counted. Pasted as part of a FileConcat bundle, each file adds its tag and the bundle a short header. We did not count it; it is small against thousands of tokens per PDF.
• Our text has page markers. The extracted text carries one # Page N line per page, as the bundle does. It is counted, so the text side is slightly higher than the words alone.

Everything but the API credential is in the repository. Counting tokens is free; it needs an Anthropic API key in an env file.

git clone https://github.com/CeamKrier/file-concat
cd file-concat && pnpm install
python3 packages/cli/scripts/build-cc-corpus.py corpus --crawl CC-MAIN-2026-34
packages/cli/scripts/run-corpus.sh . corpus out
node --env-file=<your env file> packages/cli/scripts/measure-pdf-tokens.mjs \
  corpus/pdf out/pdf apps/web/src/data/pdf-tokens.json

The first command needs DuckDB and draws every format; only pdf/ is counted here. The crawl and the seed are this sample's defaults. How many index parts our draw read was not recorded, so a rerun can pick a different 195. The last one writes the JSON this page is drawn from: the four groups, the 100-page cut, the picture counts, the rejected file, and the per-file ratios behind the figure.

How many tokens is a PDF page in Claude?

On the 1,807 text-layer pages measured here, the median page counted 2,390 input tokens as a PDF document block on claude-opus-5-5. About 1,600 of that is the page's image and the rest its text. The same page pasted as extracted text counted 856. Anthropic's documentation gives 1,500 to 3,000 tokens of text per page, which is higher than what these web PDFs held.

Is it cheaper to paste PDF text into Claude than to upload the PDF?

Almost always. 163 of 165 PDFs with a text layer cost less as text, and the median file cost 2.81 times as much uploaded. The two exceptions were single pages of very dense small print.

Do scanned PDFs cost more tokens?

They cost less per page as a PDF, about 1,586 tokens, because there is no text layer to bill. There is no cheaper text to paste, though: the text has to be recognised from the image first.

Does Claude read the images in a PDF?

Through the API, yes: each page is sent as an image along with its text, and the image is most of the cost. In the claude.ai apps a PDF of more than 100 pages is read as text only, according to Anthropic's help page.

How do I count the tokens in my own PDF?

Drop it into FileConcat and the result screen shows the token count of the extracted text. For the PDF side, Anthropic's token counting endpoint takes the PDF as a base64 document block and is free to call.

• What gets lost converting documents to text, what the text side leaves out
• Markitdown vs Tika, how FileConcat's PDF reader scores on the same 195 PDFs
• Claude file upload limit, the per-chat file cap and the PDF page limits
• Token estimation, how the count on the result screen is produced
• Formats, what FileConcat reads from a PDF