Under the hood (just enough)
Tokens
A token is the unit the model reads and writes: roughly three-quarters of an English word, about four characters. Everything is measured in tokens: your message, the hidden preamble, the reply, every tool result.
Rules of thumb
| Amount | Roughly |
|---|---|
| 1 token | ¾ of a word, about 4 characters |
| 1,000 tokens | 750 words, about 1.5 pages |
| 100,000 tokens | A 150-page document |
| 1,000,000 tokens | About 1,500 pages, if nothing else were in the prompt (something always is) |
What tokenizes less efficiently
Numbers and tables, code, and non-English text all cost more tokens per word than English prose. The ¾-word rule is an English-prose average.