How many tokens are in one PDF page?
A practical tokens-per-page range for A4 and letter PDFs, plus the reasons real documents vary.
A normal A4 or letter-size PDF page often contains about 300 to 700 tokens. If you need one planning number, use 500 tokens per dense text page. It is useful for a rough estimate, but it is not a conversion rule. When I measured four real PDFs, the median page ranged from about 280 to 1,400 tokens.
Why one page has no fixed token count
A page can hold a title and two sentences or a compact legal clause with a table. Language matters too. Punctuation, numbers, code, OCR errors, headers, and repeated footers all change tokenization even when two pages look equally full.
Image-only pages contribute no text tokens until OCR turns the image into text. A noisy OCR result can then produce more tokens than a clean digital page.
A practical range
- Light page with headings: about 150 to 350 tokens.
- Ordinary report page: about 350 to 600 tokens.
- Dense legal, academic, or technical page: 600 to 1,000 tokens or more.
- Page that is mostly a table: can pass 5,000 tokens.
Multiply the page count by 500 for a first estimate. Then measure the actual file before choosing a context window or calculating API cost.
What I measured
To test that rule, I ran four public PDFs through the same extraction and tokenizer the Token Counter uses (pdf.js and o200k_base) on October 5, 2026. The median counts only pages that have text.
| Document | Pages | Tokens | Median page | Heaviest page |
|---|---|---|---|---|
| Research paper (arXiv) | 15 | 10,621 | 816 | 1,048 |
| NIST AI Risk Management Framework | 48 | 21,419 | 501 | 628 |
| IRS Form 1040 instructions | 126 | 245,481 | 1,413 | 6,216 |
| UN Universal Declaration of Human Rights | 8 | 2,190 | 276 | 324 |
The 500 rule held almost exactly for the NIST framework, a typical prose report. It undercounted the research paper by about 60% and the tax instructions by almost three times, because references, equations, and tables pack far more tokens onto a page. The heaviest IRS page, a tax table, came in at 6,216 tokens on its own. The declaration, with short articles and generous spacing, landed well under the rule.
Two more things showed up. Averages hide the extremes: the IRS file averages 1,948 tokens per page because a handful of table pages pull it up, which is why the Token Counter now shows a per-page chart. And on all four English files, the older cl100k_base and newer o200k_base encodings landed within 1% of each other. For other languages the gap is much larger, which I measured separately in the same PDF in four languages.
Measure the PDF, not the page count
The PDF Token Counter extracts the text locally and applies the GPT tokenizer in your browser. It also shows words and characters, making it easier to spot a scan with no usable text layer.
If the result is too large for your workflow, use the PDF Chunker rather than guessing where to split. Keep room for your prompt, conversation history, and the model’s answer.