“On the ExploitBench cybersecurity benchmark, Anthropic's Mythos-5 and Fable-5 models scored 78%, compared to 34% for GPT-5.5.”