Sakana AI released Fugu-Cyber: orchestration model with 86.9% on CyberGym and 72.1% on CTI-REALM

Sakana AI released Fugu-Cyber (fugu-cyber-v1.0), a cybersecurity-specialized addition to its Fugu orchestration family. It is a third endpoint on the Fugu orchestrator, tuned for security reasoning. Sakana reports 86.9% on CyberGym (UC Berkeley benchmark of 1,507 real-world vulnerabilities) and 72.1% on CTI-REALM (Microsoft's detection-engineering benchmark). These scores surpass GPT-5.5-Cyber's 85.6% and Claude Mythos Preview's 83.1% on CyberGym, though still only incremental improvements. The orchestration works by having Fugu, a language model, build an agentic scaffold on the fly and delegate sub-tasks to specialist models. Access is gated: requires manual approval, updated Acceptable Usage Policy prohibiting offensive misuse, Token Plan only, no EU/EEA, no weights. Pricing is $6 per million input tokens, $36 output, $0.60 cached input, all 1.2× the Fugu-Ultra rate.