Sakana AI has released Fugu-Cyber, a cybersecurity-specialized endpoint for its Fugu orchestrator family, reporting an 86.9% success rate on the CyberGym benchmark and 72.1% on CTI-REALM.

  • Fugu-Cyber is not a new frontier model but a third endpoint tuned for security reasoning within the existing Fugu orchestrator architecture.
  • The model achieves 86.9% on CyberGym, slightly surpassing reported scores of 85.6% for GPT-5.5-Cyber and 83.1% for Claude Mythos Preview.
  • On CTI-REALM, Fugu-Cyber scored 72.1%, outperforming Microsoft's top Claude configurations which ranged from 0.624 to 0.685.
  • Access is gated behind manual approval, a defensive-use acceptable usage policy, and the Token Plan, with no availability in the EU or EEA.

The release positions Fugu-Cyber as a specialized tool for security reasoning tasks, leveraging the orchestrator's ability to delegate sub-tasks to specialist models.