“Sam Altman believes static benchmark scores are becoming less useful for evaluating AI model capabilities because they are heavily gamed by developers.”