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