Interestingly, when opening this page, the first thought I had was not that the benchmarks should be high, but 'I really hope they did not benchmaxx'. I think a model with modest benchmark scores can have much better real world utility as opposed to the current frontiers that are RL'd into being robotic and rigid.