A token is close to a piece of a word, not quite a character and not quite a whole word. “Running” might be one token, while a strange word, code snippet, or punctuation-heavy string can break into several, and different models tokenize a little differently.
Think about a grocery checkout. The cashier doesn’t charge by the letters on each package, they scan items, some tiny, some bundled. Tokens are the AI version of those countable units.
How it shows up
Tokens are how the model measures what it reads and writes. Your context window is counted in tokens, pricing is often based on input and output tokens, and limits and speed connect back to token count. Paste a ten-page PDF, a messy transcript, and five old email threads into one prompt and the model spends tokens reading all of it, then spends more writing a long report back. Old messages carried forward keep counting too.
Why you care
Cleaner context saves money and attention. A short, specific prompt with the right source beats a giant pile of maybe-relevant text, which is why prompt caching helps systems that reuse long instructions, and why vague, sprawling questions hit a rate limit faster. Don’t obsess over exact math. Just remember text isn’t free, context is finite, and clean inputs make AI work better.