It would be interesting to have benchmarks that take this into account (maybe they already do or I’m misunderstanding how those benchmarks work). I.e. when comparing quality between two different models of vastly different performance, you could be doing best-of-n in the faster model.