A lot of inference providers for open models only accept prepaid payments, and managing multiple of those accounts is kind of cumbersome. I could limit myself to a smaller set of providers, but then I'm probably overpaying by more than the 5.5% fee

If you're only using flagship model providers then openrouter's value add is a lot more limited