Sakana AI 推出多代理協調模型 Fugu 與 Fugu Ultra

Tokyo-based AI unicorn Sakana AI is betting that orchestration, rather than a single large model, can deliver frontier-level performance. Its Sakana Fugu system, made generally available on June 22, 2026, dynamically assigns work to multiple models and agents through one interface. Fugu-Cyber applies that architecture to security tasks, including validating vulnerabilities in complex codebases and turning raw threat intelligence into operational detection rules.

Sakana AI released Fugu-Cyber on July 21, 2026, as a pay-as-you-go option. In company-run tests, it posted an 86.9% success rate on CyberGym and 72.1% on CTI-REALM, scores Sakana said were comparable with OpenAI’s GPT-5.5-Cyber and Anthropic’s Mythos Preview. The claims have not been independently verified. Standard pricing is $6 per million input tokens and $36 per million output tokens, rising to $12 and $54 for contexts above 272,000 tokens.