What is a token, and why am I being charged for them?

A token is a piece of a word

Models do not read letters or words. They read tokens: common chunks of text that the model was taught to recognise. A short everyday word is usually one token. A longer or unusual word gets split into several.

In English the rule of thumb is about four characters to a token, or roughly three-quarters of a word. So a hundred words is around 130 tokens, a page of A4 is around 500, and a long email thread can easily be several thousand.

Names, code, unusual spellings and other languages break the rule of thumb. It is a guide for estimating, not a measurement.

DOT / THE TRANSLATOR

It is a daft unit and nobody would have chosen it for humans. It exists because it is how the model sees text, and the billing simply inherited it.

In and out are priced separately

Two numbers matter, and they are not the same. Input tokens are everything you send: your question, plus any document you attach, plus — and this is the part that surprises people — the entire conversation so far.

Output tokens are what it writes back. Output is typically several times dearer than input, because generating text takes far more work than reading it.

That asymmetry is why “be brief” is not merely a style preference. A request that produces four paragraphs instead of one costs about four times as much to answer.

What that means in actual money

Prices are quoted per million tokens, which is a unit designed for people buying in bulk and hopeless for judging whether your own afternoon will cost anything.

Translated: at a price of a few pounds per million output tokens, an ordinary back-and-forth — a few hundred words in, a few hundred out — costs a fraction of a penny. You would have to work quite hard to spend a pound in an afternoon of ordinary chatting.

The costs that actually bite are the ones that repeat or scale. Summarising two hundred documents. An agent looping for twenty minutes. A feature in your product that runs every time a customer clicks something. That is where per-token pricing stops being trivia and starts being a budget.

We do not print current prices here, because they change and a stale price is worse than none. Each model's recorded price, with the date and the source, is on its card in Orbit.

Why the same question can cost wildly different amounts

The model you picked is the biggest factor. Larger models cost many times more per token than smaller ones, and for a lot of everyday jobs the smaller one is perfectly good.

How much you attached is next. Dropping in a fifty-page PDF and asking one short question is an expensive short question.

Then there are reasoning models, which produce a long internal working-out before answering. You are billed for that thinking even though you never see it, which is why they cost more and answer more slowly.

And on a subscription, none of this appears as money at all. It appears as limits: messages per hour, files per day, a slower model once you have used the fast one. The meter is still running; you are just reading it in a different unit.

BEEP / THE MYTH-BUSTER

“Unlimited” is doing a great deal of heavy lifting in this industry. There is always a limit. Sometimes it is money, sometimes it is a message cap, and sometimes it is being quietly moved to a cheaper model at teatime.

See it for yourself

Reading about it only gets you so far. This one is written to show you the thing described above, rather than describe it again.

If you remember nothing else

• A token is about three-quarters of a word in English.
• You pay for tokens in and tokens out, and output usually costs several times more.
• Every message re-sends the whole conversation, so long chats cost more per question.
• Ordinary personal use costs pennies; repeated and automated use is what needs a budget.
• On a subscription the cost shows up as limits rather than money.