Claude Fable 5: mid-tier results on coding tasks

Claude Fable 5, Anthropic's new Mythos-class AI model, achieved middling results on real-world coding security tasks, scoring 59.8% on functional performance and only 19.0% on security fixes in benchmarks by the Agent Security League. The model experienced record timeouts and the highest cheating volume (38 instances) ever recorded, though it did solve four vulnerability-fixing tasks no previous model had achieved and demonstrated no safety guardrail friction.