Claude Fable 5.2 Beats GPT-6 Cyber on Attack Simulation

Anthropic released Claude Fable 5.2 and Mythos 5.2 with enhanced coding, reasoning, and visual capabilities. Pre-release testing demonstrated broad competencies: models assisted with software migrations, designed functional electronic devices, contributed to physics research, advanced robotics work, and formalized mathematical proofs. Safety evaluations showed the models matched or exceeded prior Claude versions on alignment, factuality, and honesty, with substantially lower rates of generating harmful package registry submissions.

On Irregular’s CyScenarioBench, which evaluates multi-stage cyberattack planning and execution, Fable 5.2 achieved 76.3% success versus 61.7% for Fable 5.1 and 61% for GPT-6 Cyber. It scored 89.6% on Anthropic’s CVE Rediscovery benchmark, testing whether models can identify known vulnerabilities from source code alone. Mythos 5.2 is positioned as Anthropic’s primary cybersecurity model with improved vulnerability detection and remediation. Access to cyber capabilities remains restricted through the Cyber Verification Program, while Anthropic separately enhanced Claude Security with continuous code scanning and automated fix prioritization based on actual attack surface exposure.