The unit behind every AI bill
Every AI bill is counted in tokens, so a token is the unit your costs are really made of. The price per token is easy to look up; what decides the bill is how the tokens add up in a real request, and which of them you’re paying full price for. Once you can see that, you can price a feature before you build it, cut its cost without losing quality, and compare models on the number that matters.
What a token actually is
A token is a chunk of text, usually about three or four characters of English. Before a model reads your words, it breaks them into these pieces: common short words become a single token, longer or rarer ones split into a few, and spaces and punctuation count too. A rough guide is that 1,000 tokens is about 750 words. The model charges by the token because that count reflects the actual work it does, both reading your request and writing its reply.
It costs more in some languages than others
The pieces are based mostly on English, so other languages need more tokens to say the same thing. Swedish compounds, German technical terms, and any non-Latin script can run 20 to 50 per cent higher for the same meaning. If your product serves several languages, the same feature quietly costs more in some markets than others, which is worth knowing before you set one price for everyone.
Why the reply costs more than the request
Every provider charges separately for the tokens you send in and the tokens the model sends back, and the reply is usually three to five times more expensive. The reason is simple: the model reads your whole request at once, but it writes its answer one token at a time, each one depending on the last. So the length of the answer drives the bill harder than the length of the question. Asking for a shorter answer is a real way to cut costs, and it often makes the feature better as well.
A real request is bigger than it looks
The text you send is rarely just the user’s message. It usually stacks up: the standing instructions you give the model, the conversation so far, any documents you look up to help it answer, and finally what the user typed. In a chat, the whole history gets sent again with every message, so a long conversation costs several times what its first line did. Looking things up adds to it too, since dropping a few documents into a request can add thousands of tokens to a single answer.
The discount most people miss
Most of the big providers now charge less for text they’ve seen recently, which in practice means the parts of your request that repeat: your standing instructions, the same reference documents, the boilerplate. The saving is large, often most of the input cost. Getting it is a matter of order. Keep the parts that stay the same at the front and identical each time, and put the parts that change at the end. Rearranging costs nothing, and the saving comes back on every request.
How to price a feature before you build it
You can get a good estimate in about half an hour:
Write three or four realistic examples of what you’d send, including your standing instructions and any documents.
Run them through your provider’s token counter. Every major provider has one, and most report the count with every reply.
Note the input and output numbers, apply the prices from the pricing page, and multiply by how many requests you expect in a month.
As a rough example, 400 tokens in and 150 out per request, at typical prices, comes to about €0.003 a request, which is roughly €300 a month at 100,000 requests. The sum is small. Doing it before you launch is the difference between choosing your price and finding it out from the invoice.
Why the same feature costs more on one model than another
Three things change at once when you switch models: how the text gets split into tokens, the price per token, and how wordy the answers are. A price sheet only shows the middle one. The honest way to compare is to run your own examples through each model and look at the cost of a whole request, start to finish, which the half-hour estimate already sets you up to do.
An AI bill is really a count of the work you asked the model to do. Every habit here comes down to asking for a little less of that work without giving up anything that matters.
