Works much worse since gpt 5.4 and codex performs remarkably better with the same model

Works much worse since gpt 5.4 and codex performs remarkably better with the same model #2632