You pay by the token

Somewhere right now, an engineer is staring at a bill that says their fun little AI feature cost $14,000 last month, and they have no idea why. This course is how you avoid becoming that engineer. And it starts with the meter itself — because AI isn't billed per request, per user, or per month. It's billed per token.

You've met tokens before if you took the How ChatGPT Actually Works course: the word-chunks a model actually reads and writes, roughly ¾ of a word each in English. Providers price them per million, with two separate meters — one for the tokens you send (input), one for the tokens the model generates (output).

Scroll to watch one request get metered.

The two meters

The pricing table for any model is two numbers, dollars per million tokens:

• Input — what you send. Cheap-ish per token, but there tends to be a lot of it: system prompts, documents, conversation history.
• Output — what the model writes. Typically several times the input price (~5× is common), because each generated token is a full forward pass through the network.

Why small numbers become big bills

A single call costing $0.0125 sounds like nothing. But token costs scale with everything: more users, longer conversations, bigger documents, chattier replies. A product with a million requests a month at a cent each is $10,000 a month — and most teams discover this after launch, not before.

The good news: the same arithmetic that makes bills explode makes savings compound. Cut average tokens per request by half and the bill halves everywhere, forever. The rest of the course is about where those cuts hide — starting with the sneakiest cost of all: the conversation itself.