Token

Ask AI how many R's are in "strawberry" and watch it fumble.

Ask enough models and one of them will look you dead in the eye and swear there are two. The smartest tool you've ever used, tripped up by a word a five-year-old can spell. Once you understand tokens, that stops being a mystery and starts being obvious.

Here's the thing: the AI never actually sees your sentence the way you do. Before it reads a single thing you typed, your words get chopped into chunks called tokens. A token is sometimes a whole word, sometimes a piece of one. "Cat" is one token. "Strawberry" gets broken into a few. Rough rule of thumb:

• One token is about three-quarters of a word.
• So 100 tokens is roughly 75 words.
• And 1 million tokens is around 750,000 words, or about 1,500 pages.

Think of them as the LEGO bricks of language.

You handed the AI a sentence. It sees a pile of little bricks, and its entire job is predicting which brick most likely snaps on next. It's not thinking about your meaning. It's playing the world's most sophisticated game of "what comes after this."

And that one fact is hiding behind three things you've probably bumped into.

1. The bill. If you use AI through an API (the drive-thru window from earlier), you don't pay a flat monthly fee. You pay by the token, both the ones you send in and the ones it sends back. A short question is cheap. A giant document pasted into a long, rambling conversation is not. Every brick has a price tag.

2. The room. Remember the context window, that one room everything has to fit inside? That room isn't measured in pages or messages. It's measured in tokens. When you hear a model "has a 200K context window," that's 200,000 bricks it can hold at once. Fill the room with bricks and the early ones get shoved out the door.

3. The blind spot. Back to strawberry. The AI isn't looking at the letters s-t-r-a-w-b-e-r-r-y. It's looking at two or three bricks that, stacked together, mean "strawberry" to it. The individual letters got swallowed when the word became bricks. Counting R's is a letter question, and there are no letters left to count.

So, practically:

• Tighten your prompts. Rambling costs tokens and clutters the room. Say what you mean.
• Don't trust AI on spelling, character counts, or "how many letters." Those are questions about individual letters, and the AI only ever sees bricks, never the letters tucked inside them. Verify those yourself.
• Convert files before pasting. A messy PDF burns a pile of tokens just to decode the layout. Plain text spends those bricks on the actual thinking instead.

The bricks aren't just what you pay for. They're what the AI is thinking with. Hand it a cleaner pile and you get a cheaper bill and a sharper answer, from the exact same model.