What are tokens?

At a glance

• A token is not a word: 100 tokens is about 75 English words, or roughly 4 characters each.^[1] A one-page document runs ~500 to 800 tokens.
• Vendors bill per million tokens, charging separately for input (what you send) and output (the reply), with output usually pricier.^[2]
• A model can only “see” a fixed number of tokens at once, the context window.^[4] Instructions, documents, and chat history all share it.
• Non-English text and code use 20 to 30 percent more tokens, raising the bill.

How it works

Your text is sliced into common character chunks before the model reads it.^[5] Frequent words like “the” are one token; a rarer word like “tokenization” splits into “token” and “ization”. Spaces and punctuation count too. You never count by hand: free online tokenizer tools count any text exactly.^[1]

Why it matters

Tokens are the meter. In 2026, prices ran from about 10 cents per million for budget models to $30+ for top reasoning models.^[3] One request costs a fraction of a cent, but across thousands of daily users that becomes thousands of dollars a month. Long chat histories get re-sent every turn, so the meter runs faster than expected.

The context window

This is the model’s short-term memory, a hard ceiling on tokens. Modern windows hold hundreds of thousands of tokens, but when you hit the limit the oldest material drops, so the AI “forgets” the start of a long chat or misses details in a big document.

Bottom line

Once you know 100 tokens is about 75 words, that input and output are billed separately, and that the context window caps what fits, AI pricing becomes a number you can estimate and control.