Qwen 3.5 27B just one-shotted something I've been fighting with Opus 4.6 about for a month. Opus just doesn't get it.
(this is several things in the category of "absolutely, categorically, mathematically, provably wrong" category). Also something its training data sees a lot.