[mauricescheff]

I love GPT 5.4, and I'm under the impression that many claims of it's good or it's bad are only based on a specific setup + a few shots. I think many model's power and not just GPT 5.4, comes from good setup + good prompting, and some folks just do like "write a game of a spaceship shooting asteroids" and then decide if it's good or not based on that. Full-time GPT 5.4 usage is expensive though, I'm averaging $300-400/month using it all day every day with one or two agents in parallel. After GPT 5.2, I think it was in December last year, the thing that changes is that it went from sometimes getting it right, to almost always getting it right, and I found the times it doesn't get it right, it's because of your context + setup + prompt and not the model itself. I haven't used claude's agents, but I would guess that they're as powerful as GPT 5.4 if used right. Maybe slightly / marginally better or worse but not much.