QUICK ANSWER
PDF Token Counter
A PDF’s token count cannot be determined reliably from its page count. The total depends on the selected AI model and on how the provider processes extracted text, page images, charts, and scanned content. For an API request, count the actual PDF with the provider’s token-counting endpoint before sending it for generation.
Why pages do not map to a fixed token total
Two PDFs with the same number of pages can contain very different amounts of selectable text, images, tables, diagrams, or blank space. A scanned document may contain page images instead of an embedded text layer. These differences change what the model receives and make a pages-to-tokens conversion unreliable.
Extracted text is only part of some PDF requests
OpenAI documents that supported vision models process both extracted PDF text and page images. The image detail setting can also change the visual input processed by the model. A tokenizer applied only to copied text therefore does not reproduce the complete PDF request count.
How to obtain the relevant count
Use the same provider, model, PDF, accompanying prompt, and processing options intended for the real request. OpenAI’s input-token endpoint accepts supported PDF file inputs and returns the full processed input count. Anthropic’s message token-counting endpoint supports base64-encoded PDFs and returns an input estimate before generation.
Which result should be used?
- For an OpenAI PDF request, follow the OpenAI token-counting guide.
- For a Claude PDF request, follow Anthropic’s token-counting guide.
- If you extracted plain text yourself, use the OpenAI Token Counter for that text only, then use the Token Cost Kit calculator for cost planning.