But that is my point: if benchmaxxing was all the labs were doing, then surely the dumber model could/would have equivalent performance? Rather than noticeably

But that is my point: if benchmaxxing was all the labs were doing, then surely the dumber model could/would have equivalent performance? Rather than noticeably worse perf on a (somewhat trivial to game) test.