I think Codex is probably marginally stronger than Opus in my testing.

But it's much much worse at writing issues than Claude models.