A token is a chunk of text, about four characters, and it is the unit a language model reads and writes in. Everything you send an LLM and everything it sends back gets counted, and billed, in tokens.
Tokens get called the currency of AI, which is fair on the billing side.
I like to call it the alphabet.
Luka Eric, 0:05
Letters are the units English is built out of. Tokens are the units an LLM is built out of, and there is nothing smaller for it to work with.
How does tokenization work?
Take the sentence “Customers prefer subscriptions and emoji over one-time purchases.” The split is not the one you would make. “Customers” is one token. “Prefer” is one token. “Subscriptions” is three. The emoji is a token on its own.
The splitting is called tokenization, and every model ships its own tokenizer, so the same sentence comes out at slightly different counts on GPT and on Claude.
Those cut points have nothing to do with meaning. A tokenizer learns the most common chunks of text in the material it was trained on and cuts there, so everyday words survive whole and rarer ones get sliced. “Unbelievable” comes apart into three pieces, un-believ-able. That is the shape to carry around: common word intact, unusual word in fragments, and no relationship at all to where you would put a hyphen.
Is a token the same as a word?
No. Short common words are one token each, longer and rarer ones break into several, and spaces, punctuation and emoji are all counted too. Over normal English prose it averages out to about 100 tokens per 75 words, which is the ratio to budget with.
How many words is 100 tokens?
About 75. Every other conversion comes off the same two ratios: one token is roughly four characters, 100 tokens is roughly 75 words.
| Text | Tokens |
|---|---|
| A 280-character tweet | about 70 |
| 75 words | about 100 |
| A 150-word email | about 200 |
| One page of a document | about 700 |
| 1,000 words | about 1,300 |
| A ten-page document | nearly 7,000 |
| A 300-page manual | about 210,000 |
Paste a ten-page maintenance contract into a chat and you have spent nearly 7,000 tokens before the model has written a single word back.
Why a long chat costs more than a short one
Input tokens, the ones you send, are cheap for a model to process. Generating the reply is the expensive part, so providers charge roughly three to five times more for output tokens. As of July 2026, Claude Opus 4.8 lists at $5 per million input tokens and $25 per million output, and Fable 5 at $10 in and $50 out. So when you cost a workflow, the length of the answer matters more than the length of the question. An agent that drafts a full quote for a solar install costs more to run than the same agent answering whether the job is in your service area.
Every turn of a chat carries the earlier turns back in with it. We covered context windows in the previous lesson, and the short version is this. Turn one might be 100 tokens of input. The answer comes back, you ask a second question, and the model is now reading 400 tokens, because your first question and its answer came along for the ride. By turn ten, that same conversation is a 10,000 token context window. You pay for the start of the chat again on every turn.
How do I use fewer tokens?
A bloated prompt carries information the job does not need, and you pay for that dead weight on every turn it stays in the window. A focused prompt names one outcome and gives only what serves it: who, what, when, one worked example, and the criteria you will judge the result against, down to “don’t go over 300 words”. The full recipe is in the prompt lesson. For an after-hours intake agent it means the caller and the property, the details you want captured, when to escalate to the on-call tech, and a word limit. Fewer tokens is cheaper, and it also gets you a better answer.
