AI model context windows
200k tokens doesn't mean 200k words
Most operators assume a 200k token context window holds 200k words—the real math tells a different story, and it matters for long-form work.
The cursor blinks at the bottom of your 47-page style guide, ready to paste the whole thing into Claude. The interface says “200k token context window”—enough room for roughly 150,000 words, right? Wrong. Most English words map to one or two tokens, but technical terms, URLs, formatting markup, and non-English strings inflate token counts fast. That 47-page PDF might be 60,000 words but cost you 95,000 tokens once you account for markdown formatting, code snippets, and brand names spelled in camelCase.
Today’s edition walks through the token-to-word ratio most operators get wrong, why it matters when you’re feeding long documents into AI models, and how to estimate token costs before you hit a context limit mid-task.
In today's email:
- How tokens and words differ—the encoding that trips up long-form operators
- Why markdown, code, and URLs inflate token counts faster than plain text
- Calculating token costs before you paste 60,000 words into an AI model
- When to split documents instead of relying on extended context windows