Every new model reports 90%+ on the same benchmarks. Why don't those scores help us choose, and what should we use instead?

Make your call first. Use the published answer and rubric to examine your reasoning when you are ready.