That's a wild statement to me. Even with spending significant time making plans with Opus 4.7 and GPT 5.5 on xhigh, I still find lots of poor decisions made when it actually goes to implement it. I find the quality of PRs hasn't dramatically changed either way because the better engineers will spot the issues whereas others will find what the AI is doing acceptable.