You are selection biasing towards the most extreme cases of AI absurdity.

GPT-5.3-Codex and Opus 2.6 were released. Reviewers note they're struggling to find tasks the previous versions couldn't handle. The improvements are incremental at best.

I have not seen any claims of this other than Opus 4.6 being weirdly token-hungry.