Sakana AI Launches Multi-Agent System Fugu: Outperforms GPT5.4 and Opus4.6 in Benchmarks

According to Golden Finance, Japanese AI development company Sakana AI has announced the release of its flagship commercial product, Sakana Fugu, and has opened applications for early Beta testing. The system is positioned as a multi-agent orchestration system, provided in the form of a single-model API compatible with OpenAI formats. It includes the low-latency-optimized Sakana Fugu Mini and the task-intensive Sakana Fugu Ultra. The product architecture is based on the team's Trinity and Conductor papers published at ICLR 2026. At its core is a lightweight, self-learning language model that does not rely on manually preset team roles or fixed pipelines. Instead, it dynamically invokes a pool of Worker models and allocates tasks based on task complexity. The system also supports scaling during testing and features adaptive recursive invocation capabilities. The model can read its previous outputs as context, enabling it to autonomously identify flaws during operation and initiate error-correction workflows. Users can adjust the recursion depth as a computational axis during inference. Evaluation data shows that Sakana Fugu Ultra demonstrates performance surpassing cutting-edge single models in hardcore reasoning and coding benchmarks. In the GPQAD, LCBv6, and SWEPro tests, Sakana Fugu Ultra achieved scores of 95.1, 93.2, and 54.2, respectively, outperforming mainstream flagship single models such as GPT 5.4, Gemini 3.1, and Opus 4.6.