Sakana AI has launched "Sakana Fugu," a service that achieves high performance by dynamically combining multiple powerful models. The service eliminates dependence on a single vendor and handles complex multi-step tasks. Users can access a group of specialized models through a single API, with Fugu automatically handling the selection and switching of models for each task. The company stated that this reduces API-related complexity while improving cost-performance.
Sakana AI Launches Sakana Fugu, a New Service for Dynamic Orchestration of Multiple LLMs
This article is a translation. Read the Japanese original
The technical foundation lies in two papers presented at ICLR 2026: "TRINITY" and "Conductor." TRINITY utilizes a lightweight evolutionary coordinator to oversee multiple LLMs, adaptively distributing work by assigning roles such as "Thinker," "Worker," and "Verifier" to each model. Conductor uses reinforcement learning to autonomously design natural language-based coordination strategies and prompts between agents. According to the company, this allows a diverse collection of LLMs to exhibit capabilities that surpass those of any single model.
Three types of models are offered: "Fugu," "Fugu Ultra," and "Fugu Cyber." All are available via an OpenAI-compatible API. "Fugu" balances high performance with low latency and is intended for standard integration into daily coding tasks and chatbots. Users can also exclude specific agents from the model pool to meet data privacy and compliance requirements. Additionally, the company is working toward compliance with EU/EEA regulations such as GDPR, and the service is currently not available in those regions.
Sources: Sakana Fugu (HN 247pt, 127 comments) (HN Search (backfill), 2026-06-22)