Tokens, and why they decide the price

Models do not read letters or words. They read tokens — chunks of text, usually a few characters each.

Almost every practical limit and every price you will encounter is denominated in them, so this is worth ten minutes.

What a token is

Roughly: three-quarters of a word in English.

• `cat` is one token
• `running` is likely two: `run` + `ning`
• `unbelievable` is several
• Common words are single tokens; unusual words, names and technical terms get split

The splitting is learned from data rather than designed by hand, which is why it looks arbitrary. It is also why the model can struggle to count letters — it never sees letters.

The numbers worth memorising

You do not need precision. You need to be able to tell whether something is hundreds or hundreds of thousands, because that is the difference between a rounding error and a bill.

Why this matters three times over

Limits. A context window of 200,000 tokens sounds enormous — and is, at roughly 150,000 words. But a long conversation plus a pasted document uses it faster than people expect.

Cost. Every provider prices per million tokens, in and out. See what it actually costs.

Speed. More tokens means more time, roughly linearly. A 50,000-token document takes far longer to read than a question.

Input and output are not the same price

On nearly every model, output costs several times more than input — often three to five times.

This is why a long, chatty answer is expensive in a way a long question is not, and why "be brief" is a cost-control instruction as much as a stylistic one.

The thing that catches people out

In a conversation, everything is re-sent every turn — see why it forgets. So a long thread is not billed once per message. Each new message re-processes the entire history before it.

The practical consequence: long conversations are disproportionately expensive. Starting a fresh thread for a new task is the cheapest optimisation available.

Why non-English costs more

Tokenisers are trained mostly on English, so English is compressed efficiently — common words are single tokens. Many other languages need more tokens for the same meaning, sometimes two or three times as many.

A paragraph in English and its translation into a language with a less efficient tokeniser are billed differently, for identical content. Systems that quote per word rather than per token are quietly hiding this.

Why you should care even if you never pay

Free tiers have token limits rather than bills. The limits bind in exactly the same places: long documents, long conversations, and pasted material.

The arithmetic is the same. Only the currency differs.

Where to go next

• Why it forgets
• What it actually costs
• What Q4 actually costs you — the same trade in memory rather than money