Unit Cost Desk

Notes on the unglamorous half of running a generative product: what one output costs to serve, who absorbs it when the bill moves, and how to price it without inventing numbers.

Most writing about image models is about the pictures. Almost none of it is about the invoice. That gap is where the avoidable mistakes live — a free tier with no ceiling, a credit that costs three times more to serve in one cell of the grid than another, a comparison page that loses trust because it claims to win everywhere.

These come out of building a browser generator for GPT Image 2.5 and having to answer each of them with a number rather than a guess.

Notes

• A daily spend ceiling that degrades the right thing If the ceiling refuses paid work too, you will raise it until it does nothing. The fix is asking which pool paid.
• Two models, one token rate When the quality model bills identically to the fast one, "premium for quality" is a price with no cost behind it.
• The comparison page that names the other side's wins first Why the page that concedes gets quoted and the page that sweeps does not.