Somebody on your team has already priced the feature. They opened the provider's pricing page, multiplied tokens by requests, added headroom, and came back with a number.

That number isn't wrong. It is the correct price of the thing they costed, which turns out to be about a quarter of what the feature costs to run for a year.

The rest arrives on a schedule. Month two brings evaluation, because the golden set runs every time somebody edits a prompt and early on you are editing constantly. Month three brings retries, which cost exactly what successes cost, and the person reviewing low-confidence output whose time sits in another budget. Month four is the difficult one: the re-index and the first audit request land in the same week, and the traces that would have answered the audit were sampled at 1%.

None of it is hidden. It is distributed across other people's budgets and other people's months, which is why it never reached your estimate.

The infographic follows one feature through its first year, printing each cost in the month it actually appears, then separates what grows with usage from what you pay at ten users or ten million.