AI Credits

GitHub's HydraFusion router bills every model it invokes at standard rates, and two of its three patterns fire two or three models per turn. The measured saving swings from 36% to 67% across just three benchmarks — and the one that looks most like enterprise work saved least.