How many tokens is a PDF?
Search this question and you'll find a different percentage on every site, none of them reproducible. Here's the arithmetic from vendor documentation instead — and how to measure your own file in ten seconds.
Published 29 August 2026
The only well-documented numbers
The most concrete public figures come from Anthropic's PDF documentation, which describes two ways a PDF can be processed:
- As extracted text only: a 3-page PDF uses approximately 1,000 tokens — about 333 per page.
- With full visual understanding — each page processed as an image and as text, which is what lets the model read charts, layouts and figures: the same 3-page PDF uses approximately 7,000 tokens — about 2,333 per page.
That second mode is what an AI chat effectively does when you drop a PDF in and expect it to understand everything on the page. The ratio is the headline: the same words cost roughly seven times more inside a PDF than as plain text.
Scaled up: 1 page to 600
Extending those documented per-page figures linearly gives a working table. (Real pages vary — a dense table-heavy page costs more than a sparse title page — so treat these as the sensible middle, not a guarantee.)
| Pages | As PDF (page-as-image) | As plain text | Difference |
|---|---|---|---|
| 1 | ~2,300 | ~330 | ~2,000 tokens |
| 10 | ~23,000 | ~3,300 | ~20,000 tokens |
| 50 | ~117,000 | ~17,000 | ~100,000 tokens |
| 100 | ~233,000 | ~33,000 | ~200,000 tokens |
| 600 (Claude's page cap) | ~1,400,000 | ~200,000 | ~1,200,000 tokens |
Two of those rows deserve a second look. At 50 pages, a single PDF can consume more than half of a 200K context window before you ask your first question — which is why one upload so often precedes the “maximum length” message. And at 100 pages, the PDF form simply doesn't fit in a standard window at all, while the plain-text form fits several times over.
Why PDFs cost so much
A PDF page is a drawing, not a text file. To be sure of reading everything — the chart labels, the table drawn as lines, the footnote in the margin — the model looks at a rendered image of the page alongside whatever text could be extracted. Images are expensive inputs. You pay that visual surcharge on every page, including the ones that are nothing but ordinary paragraphs.
Markdown flips the deal: everything is text, so there is no image to pay for, and the structure — headings, lists, tables — survives as characters the model reads at normal prices.
Measure your own file — the honest method
Averages are averages; your document is specific. Drop it into our PDF to Markdown converter and the result card shows tokens before and after for your actual file.
Method, in the open: the counts are estimates computed as characters ÷ 4, the standard rule of thumb for English text — real tokenizers differ by model and by content, so expect the true figure to sit within about ±20% of the estimate. The ratio between before and after is what matters for deciding whether to convert, and the ratio is robust. Everything runs in your browser; the file never leaves your machine.
Rules of thumb worth remembering
- A PDF page uploaded to an AI chat: ~2,300 tokens. The same page as text: ~330.
- Multiply pages × 2,300 to guess what an upload will cost; multiply by 330 to guess the converted cost.
- Past roughly 80–90 pages, the PDF form outgrows a 200K context window on its own. Converted, the same document uses about a seventh of it.
Measure your own PDF, free
Drop your file into the converter — it runs in your browser, nothing uploads — and the result card shows estimated tokens before and after, for your document rather than anyone's average.
Open the converter →