The unit you pay in

E2·L9 covered how to choose a tier. This module is about the mechanics underneath — what you are actually buying, why the amount changes, and how to stop being surprised.

It starts with the unit, because almost every cost question resolves to it.

Tokens are chunks, not words

Models do not read letters or words. They read tokens — chunks somewhere in between.

In English a token is roughly three-quarters of a word. Common words are one token each. Longer or unusual words split: unbelievable might be three, Bhattacharya several. Spaces and punctuation count.

A useful rough figure for English: 100 tokens ≈ 75 words. A page is about 500. This book-length article you are reading is a few thousand.

You do not need precision. You need the shape, because the shape explains everything else in this module.

The part almost nobody is told

Tokenisers were built largely from English text. So English packs efficiently, and other scripts do not.

The same meaning, expressed in Hindi, Bengali, Tamil, Arabic, Thai, Korean or Amharic, frequently costs several times more tokens than in English. Non-Latin scripts fare worst; the effect is well documented and varies by language and by tokeniser.

Three consequences follow, and they are not small:

You pay more for the same work. Where billing is per token, writing in your own language costs a multiple of writing in English.

You hit limits sooner. A document that fits comfortably in English may exceed the same context window in Bengali.

Free tiers run out faster if they are metered by tokens rather than messages.

Put beside E10, this is an equity problem with a mechanism: the people already getting lower-quality answers are also paying more for them. Worth knowing, worth naming, and it is improving as tokenisers get better — unevenly, and for the same commercial reasons E10 described.

You pay for the whole context, not your question

The second thing people are surprised by.

Every time you send a message in a conversation, the model is given the entire conversation so far — it has no memory between calls, so the history is re-sent each time. Add an attached document and that goes too.

So a long thread does not cost the same each turn. It costs a little more every turn, because the context keeps growing. Message forty is substantially more expensive than message one, for the same length of question.

L2 is what to do about that. For now, the rule: cost tracks the whole context, not the thing you typed.

Input and output are priced differently

Most providers charge separately for tokens in and tokens out, and output is usually the more expensive of the two — often several times.

That makes “give me a long detailed answer” a real cost decision rather than a free preference, and it makes “answer in one paragraph” a genuine saving rather than only a readability choice.

Where this does not apply

Be clear about scope, because most readers are not on metered billing.

A flat monthly subscription is not billed per token. There you are limited by message caps or usage limits instead, and the token mechanics matter indirectly — long contexts still consume your allowance faster.

Per-token pricing applies when you use an API, most automation platforms, and some pay-as-you-go tiers. If you are building anything from the Practitioner track, that is you.

Try it now (5 minutes)

Take a paragraph you wrote. Most providers publish a tokeniser tool — paste it in and see the count.

If you work in a non-Latin script, paste the same meaning in both languages and compare. That ratio is your personal cost multiplier, and most people have never seen it.

Check your understanding

1. When a provider charges per token, one token is: One word A chunk roughly three-quarters of an English word A single character

2. The same meaning written in Hindi rather than English usually costs: The same, since meaning is identical Several times more tokens Fewer tokens, as the script is denser

3. Why does message forty in a thread cost more than message one? Providers raise prices during long sessions The whole conversation is re-sent each time, so the context keeps growing Later answers are always longer

Recap

Tokens are chunks between letters and words, roughly 100 to 75 English words — and considerably less efficient for non-Latin scripts, so the same meaning costs more in many languages. You pay for the whole context, which grows with every turn of a conversation, and output tokens usually cost more than input. Per-token billing applies to APIs and automation, not flat subscriptions.