OpenAI's GPT-5.5-Cyber Reaches 85.6% on CyberGym Benchmark, Outpaces Anthropic's Banned Mythos Model

OpenAI announced the full launch of GPT-5.5-Cyber on June 22 as part of its Daybreak cyber defense program. On the CyberGym benchmark—which tests AI agents on reproducing 1,507 known software vulnerabilities from 188 open-source projects—the model achieved 85.6%, outperforming Anthropic's Mythos 5 at 83.8% and Claude Opus 4.7 at 73.1%.

The launch occurs as Anthropic's Mythos 5 and Fable 5 remain offline following a U.S. government ban issued June 12, citing national security concerns and potential jailbreak risks. In contrast, GPT-5.5-Cyber is available only to verified security professionals and has been pre-approved by federal agencies including the Office of the National Cyber Director. OpenAI has partnered with cybersecurity firms including CrowdStrike, Cisco, and Cloudflare to integrate the model into their tools.