Charging for the retry is the worst line in AI credit pricing

Every credit system I've used bills on the API call, not the outcome. So when a model returns something plausible and wrong, the user pays, then pays again for the version they actually wanted. You've priced your own failure rate and handed the bill to the person who noticed it.

The counter I keep hearing is that inference costs money either way. True, and beside the point. If a generation is unusable because we routed it to the wrong model, that's our decision, not their prompt.

Two things fall out of eating it yourself, both good. You have to define unusable narrowly enough that it can't be gamed, ours is a same-prompt regenerate inside a short window and nothing else. And you start caring a lot about routing, because a bad route now costs you directly instead of quietly costing someone else.

Curious who's actually doing this and where it broke for you. Most pricing pages I read still bill per attempt and call it usage.