Keeping AI costs under control in a subscription app

A subscription has a fixed price and generation has a variable cost. If the second grows faster than the first, the best month in the app's history is also the one that loses the most money. These are the habits that keep our generation costs boring.

One number, weekly

Cost of generation per active subscriber per month, by app. We look at it every week next to the subscription price, and every product decision that touches generation is checked against it. If you do not know this number, you do not know whether your subscription is profitable.

The right model for each step

Writing a story, checking it against the "never mention" list, drafting an illustration prompt and narrating the text are different jobs with different quality needs. Each uses the smallest model that does its job well. The expensive model is used where it changes what the user experiences, and nowhere else.

Generate what will be used

One story per night, generated when the parent opens the app, not a library of stories nobody asked for. One picture per story. A scene when a person asks for one. Generating ahead of demand is where AI budgets quietly disappear.

Cache what repeats

Narration of a story the family replays does not need to be generated twice. Voices, styles and prompt scaffolding are reused. The marginal cost of the second listen is close to nothing, and that is where a nightly ritual becomes affordable.

Limits that feel like the product

A story a night is a feature of a bedtime app, not a quota. A fresh deck each morning is the ritual, not a cap. Where limits exist, they are designed to match how the product is meant to be used, so nobody ever meets a meter.

Watch the tail

A small number of users generate far more than average. Sometimes they are the most engaged people you have; sometimes they are scripts. Knowing which is which, and designing for the first while protecting against the second, is the last piece of keeping the number steady.