"Reasoning: GPT dominates pure math. Frontier math 47.6% versus Claude's 27.2%. That's a massive gap. But Claude leads on Humanity's Last Exam, 53.1 versus 39.8."