According to Ethan Mollick on X, Claude3 Opus drafted an academic paper and calibrated identification to 4.5 after checks; GPT5.5 Pro flagged key errors.