Charging a credit for a bad generation is the fastest way to lose a creative user
Nobody generates once. In a creative tool the real unit is six attempts and one keeper, so credit pricing quietly charges people six times for the thing they actually wanted. The arithmetic can match per output pricing exactly and it still feels worse, because every failed try has a visible price tag on it.
The version that hurts most is the near miss. An output that's obviously broken costs a credit and two seconds. An output that looks right and is subtly wrong costs a credit, plus the minute spent deciding, plus the retry. That's the one people are thinking about when they cancel.
Two things I keep going back and forth on, and we run credits on Zeplik so I've had to pick. Price the session rather than the call, so retrying inside a window is free. And let people mark an output unusable, refund that credit automatically, and treat the flag rate as your quality metric instead of something to argue with support about.
If you've run credits at any real volume, what did your refund or dispute rate actually look like, and did you keep the model.
Replies
No credits here so I can't give you a dispute rate, but the near miss point maps onto something adjacent and I think it generalises.
Our unit isn't a generation, it's a customer interview, and the bad outcome isn't a broken output either. It's a plausible one. Something that reads fine and quietly overstates what the person actually said. Nobody flags that. They just stop trusting you.
So we moved the flag upstream. The model decides whether there's anything real in the transcript before it writes, and if there isn't, it refuses and produces nothing. Same effect as your auto refund, except the system flags itself instead of waiting for the user to notice and go argue with someone about it.
The catch is that refusing has to cost you something or you'll refuse everything. Ours is that a modest but genuine result still has to pass. Without that constraint the gate stops being a quality metric and just becomes a way of never being wrong.
@venkat_a2 Refusing has to cost you something is the constraint I hadn't written down, and it breaks the auto refund idea. If flagging is free for the user then the flag rate measures how generous the button looks, not quality. Yours is the better shape anyway because it doesn't wait for someone to notice, and on a near miss they mostly don't. What I'm stuck on is what the equivalent of a modest but genuine result is for an image, because there's no transcript to check it against.