I'm an Opus stan but I'll also admit that 5.4 has gotten a lot better, especially at finding and fixing bugs. Codex doesn't seem to do as good a job at one shotting tasks from scratch.

I suppose if you are okay with a mediocre initial output that you spend more time getting into shape, Codex is comparable. I haven't exhaustively compared though.

Yes, GPT 5.4 is better at finding bugs in traditional code. This has been easy to verify since its release. Its also worse at everything else, in particular using anything recent, or not overengineering. Opus is much better at picking the right tool for the job in any non-debugging situation, which is what matters most as it has long-term consequences. It also isn't stuck in early 2024. "Docs MCPs" don't make up for knowledge in weights.