> any evidence of this.
"Any evidence" you say? I think something has moved.
As of Autumn/Winter 2025, I can say "Here's a complete enough spec of what I want cloned. Crank on it for a few hours. Give me the clone site as I specified." And frontier agentic tooling does a hands-off YOLO job really well. (Notice: I set it up with good specs/rules to scaffold.) Cost maybe an hour of my time to set up/scaffold, and 3 hours cranking on its own, on a $20 or $60/mo sub.
I think taking the same problem to an "offshore house" (or even Fiverr or whatever) would probably easily cost 10-100x more, and quite possibly with worse (less reusable or less best-practices code quality or less functional or less clean) output overall.
So I think your OVERALL point may stand. I'm just nitpicking the specific aspect quoted above, and maybe to some degree the aspect of
> "LLM generated code is similar, but arguably more expensive and lower quality"
Which to be fair obviously depends on what code one is comparing.
I used to say "nah, it can't replicate apps that well by itself", but then I tried it (on the advice of a fellow commenter here on HN), and was surprised myself.