Edits went up 3x the week we stopped charging credits for them
Our meter used to treat every change the same. Ask for a new image, one credit. Change the text on an image you already made, one credit. It reads as fair on a pricing page and it quietly taught people to accept whatever came out first, because a fix cost the same as another roll of the dice.
We split the meter earlier this year. Generation costs credits, editing something you already paid for is free. Edits per user went up about 3x in the first week and generations barely moved, so the revenue hit was much smaller than the spreadsheet predicted. What we actually bought was people finishing work instead of abandoning it at 80 percent.
The part I got wrong is that I priced by what the GPU cost me, not by what I wanted people to do. An edit is cheap compute and expensive behaviour, and the meter should have said that from day one.
If you sell credits, what have you decided not to charge for?
Replies
Dial
retries after our own system clearly messed up, not the user. if the output errors out, times out, or is obviously broken on our end, charging for that attempt just teaches people to distrust the meter entirely, they can't tell your bug from their bad prompt. the GPU cost is the same either way, but the trust cost of billing someone for your outage is way higher than the revenue. same instinct as your edit example: price for the behavior you want, not just the compute you burned.
We do refund on our own errors, and the hard part turned out to be classifying them. A timeout is obvious. A video that renders fine but the face drifts is a bad output, not a failure, and the meter can't tell the two apart. So honestly we refund what we can detect and eat the complaints on the rest. Where do you draw that line, or do you just refund anything a user flags?