More like the "gooner/jailbreak tax"

For my product I don't charge the user for blocked requests because content classifiers aren't perfect and sometimes trigger frivolously... but in theory a persistent user intentionally trying to push boundaries could cost me a lot of money: since I'm still paying for resources used to do the inference that was blocked.

This seems like an ok way to deal with that gap, but only if there's a reasonable amount of grace built in that they're not detailing here (since detailing it will result in people abusing right up until said grace expires)