Especially Claude is well known for wasting output tokens, for maximizing output token use when fewer tokens would do just fine, although other models too have picked up this disease as of late.

Yes, better models could produce better output, especially for a large project, but in my experience, the quality of the output depends 10x more on the clarity and refinement of the input. In the real world, the bulk of engineering is incremental, not one-shot.

Also, when I see a large repo with just three commits made all at once, it tells me that the vibe-coded output hasn't really been reviewed or refined over time, that it has not withstood the test of time at all, it hasn't received the love and attention it needs to make it mature, and so it cannot be trusted in this stage of its development.