@xpasky but 4o is also trained with a different regime, i think, than most of these other models (not outcome-based RL on math and code stuff), so in terms of reward hacking, it's less clear where it falls