comparing o3-pro reasoning to gemini 2.5 pro and claude 4 opus on a speculative, open-ended prompt