Claude Sonnet 4.5 has emerged as the best-performing model in ‘risky tasks’, narrowly edging out GPT-5 in early evaluations ...