Opus 4.6 was the best model in the family, following two ones were seriously brain damaged to do well on benchmarks.
replyyeah those have been horrible
replyI wouldn't say terrible-terrible but only better at from-prompt-to-solution rather than interactive discussing and problem solving.
replyI tend to define it "better at solving, worse at assisting phenomenon". Which doesn't properly show on benchmarks that only focus on the solving part.