Token cost / cost discipline

Start with the unit, because the whole subject rests on it: AI is billed by the token — a word-piece, roughly three-quarters of a word — counted both ways, on what you send in and on what comes back. A meter running in both directions.

Cost discipline is treating that meter as something you design around rather than discover on the invoice: the cheapest tier that can do the job, caching so the same question isn't paid for twice, batch for anything that can wait until tonight, a hard cap on volume — and every job that runs on a timer labelled with what it costs per month.

Why it matters: AI costs grow with use, quietly, and success is what breaks the budget — the feature everyone loves is the feature that runs up the bill. Nothing warns you. The habit that saves you is dull and it works: know what a job costs before you set it running every night.

Related: Token · Model tiers · Caching · Batch