Claude Mythos leaks, crushes Opus 4.6 benchmarks

Mythos marks the transition from LLMs that assist to models that act autonomously, specifically bridging the gap in complex cybersecurity tasks that previously required human intervention.

• –SWE-bench Verified scores in the mid-to-high 80s suggest Mythos can handle multi-file repo maintenance with minimal supervision.
• –The jump in autonomous exploit development (90%+ success on JS shells) explains Anthropic’s cautious, gate-kept preview rollout.
• –Codenamed "Capybara," the model introduces a new pricing and performance tier above the existing Opus line.
• –Terminal-Bench 2.0 scores exceeding 75% point toward a future of fully autonomous DevOps and system administration agents.