Skills
Claude Mythos Preview Leads SWE-bench Pro at 77.8% — But You Can't Use It: Restricted to Project Glasswing's 40-Company Defensive Cybersecurity Coalition
As of May 13, Claude Mythos Preview tops SWE-bench Pro at 77.8%, ahead of Claude Opus 4.7 Adaptive (64.3%) and GPT-5.5 (58.6%). SWE-bench Pro is the contamination-resistant benchmark that typically cuts frontier scores by 50-60 percentage points versus Verified. But Anthropic stated in April that Mythos Preview will not be made generally available — it powers Project Glasswing, a restricted coalition of ~40 companies (Apple, Google, Microsoft, Amazon, NVIDIA, Cisco, JPMorgan Chase) using Mythos for defensive cybersecurity. OpenAI separately stopped reporting SWE-bench Verified scores after finding 59.4% of hard test cases were flawed, recommending Pro instead.
Source
↳ Follow the thread