Claude Mythos Preview tops coding, security tests

This reads like a true capability jump, not a routine model refresh: Anthropic is showing a Claude tier that pushes harder on agentic work while staying gated behind security concerns.

• –Official Anthropic results show 77.8% on SWE-bench Pro, 82.0% on Terminal-Bench 2.0, 93.9% on SWE-bench Verified, and 83.1% on CyberGym.
• –The CyberGym result matters because it frames the model as a defender-first security tool, not just a better coding assistant.
• –Anthropic explicitly says it does not plan to make Mythos Preview generally available, which suggests safety and misuse risk still constrain deployment.
• –For developers, the practical signal is that benchmark leadership is moving fast, but access and guardrails still determine whether a model is usable in production.