Tokens aremoney.
Step an ordinary conversation forward one turn at a time and watch exactly which part of it makes the bill climb.
Watch where
the money goes.
You are not billed by the message. Every message you send is cut into tokens and charged by the piece, and so is every word that comes back. What follows is one ordinary ten-turn conversation with an assistant, running on real token counts, so you can see which side of it costs what.
One turn, in full.
Here is the first turn as the model sees it. Your message is on the left, its reply is on the right, and both have been cut into genuine tokens rather than estimated from the length. Notice that the space in front of a word usually travels inside the token, and that punctuation tends to be a token all by itself.
Honestly, that quote for rewiring the shop is unbelievable. Can you help me write a polite reply that asks them to break the price down properly?
Cheap to read, and the whole of it gets sent again on every turn that follows.
Of course. A quote that arrives as one figure is hard to question, so the polite move is to ask for the parts rather than the total. You could send something like this. Thank you for coming out and for the quote. Before we commit, please could you break the figure down into labour, materials and anything you are charging as a day rate? It would also help to know which lines are fixed and which are estimates. That keeps the tone friendly while making it clear that you expect the detail.
Every reply in this run is set to this length. Token for token this is the dearer side, and the meter below shows how the two sides compare as the conversation runs.
Loading the tokeniser so the counts on this page are real rather than estimated.
The meter.
Providers quote a price per million tokens, and they quote two of them. Pick a model and the whole page re-prices itself.
GPT-5 from OpenAI charges $1.25 for every million tokens it reads and $10.00 for every million it writes, so a written token costs 8 times what a read one does.
Each press sends the next message in the same conversation. The messages stay about the same size throughout, so anything that changes on the meter is the history, not the question.
The receipt.
One line prints per turn. Read is everything the model had to take in to answer, which is the conversation so far plus your new message. Wrote is the reply.
| Turn | Read | Wrote | Cost |
|---|---|---|---|
| 1 | — | — | — |
| Total | — | — | — |
Loading the tokeniser so the counts on this receipt are real rather than estimated.
Now multiply it.
One conversation costs little enough on its own to feel free. The bill only becomes real at volume.
Takes the conversation on the meter above, exactly as you have it set, and runs it this many times a day for thirty days.
Token counts come from OpenAI's o200k tokeniser. Other providers tokenise differently, and Anthropic's current models produce around a third more tokens for the same text, so the Claude figures here run low. A real request also adds a little framing around each message. The meter bills only the reply you can see, while these models also bill the reasoning tokens they work through before answering, so a real invoice runs higher. It bills every re-sent token at the full input rate too, where most providers now cache a repeated prompt and charge a fraction of that rate for the part they recognise. Prices are published list rates in US dollars per million tokens and they change often. Provided by the Institute of AI for interest and learning.
Why the bill
climbs.
Billed by the piece
A token is a chunk of characters rather than a word. Common words arrive whole, unusual ones break into several pieces, and a comma or a full stop is normally charged as a token of its own.
Two meters, not one
You pay for the tokens a model reads and again for the tokens it writes. Token for token, writing is the dearer of the two, several times over on every price list on this page.
The history is re-sent
A model keeps nothing between turns, so each reply re-reads the whole conversation so far. That is why the tenth turn of a chat costs so much more than the first one did.
Shorter answers, smaller models
Asking for a brief answer trims the expensive side of the bill, and moving routine work to a smaller model can change what it costs by an order of magnitude.
Membership is free. Accreditation is the standard.
Keep
exploring.
Get AI is a collection of games, experiences, and tools that make AI easier to understand from the Institute of AI.