At work I am sitting on a small pile of incomplete/wrong/missing-the-point bug reports right now, all generated by Opus 5 on High effort. Even under good conditions, LLMs are still wrong quite a lot, and confidently so. I can see why you think instant dismissal of LLM generated work is shallow, but I think it's at least as shortsighted to assume that when it creates poor quality work it must be an old/cheap model, bad settings, bad prompting, etc.

I'm not arguing that it's great. I'm arguing that the future of human work in this area will be limited to dealing with the cases when it isn't great, i.e. the pile on your desk. Choosing the best model and highest effort is the due diligence required to minimize the size of that pile.