It was certainly almost RL-fried to overfit the benchmarks, at the expense of actual usability. See Opus 5.

is it a benchmarkmaxxing model?!