I did some very quick tests asking Gpt5 and Claude Sonet 4 to build the same thing and while it's great that Gpt5 doesn't tell me I'm "absolutely right" all of the time, Claude is still the better coder. The task I assigned was simple, refactor some code that was really repetitive, the Gpt5 version actually ended up with more code that was even less readable after the refactor.