Yep same with latest Claude models - code isn't really any better than Opus 4.5/4.6, but use 5x as many tokens doing random stuff that's mostly unnecessary.

And yeah still for some reason they often can't understand how to set up any project locally without handholding, which is something you'd think an LLM would actually be good at