Sakana AI released Sakana Fugu, a new multi-agent orchestration system that functions as a single foundation model. The system dynamically delegates complex, multi-step tasks to a pool of specialized language models, managing selection, communication, and synthesis internally through an OpenAI-compatible API. This architecture builds upon the company's research into learned model coordination, specifically the Trinity and Conductor systems presented at ICLR 2026.

The product is available in two versions: Fugu, designed for high performance with low latency in tasks like coding and interactive services, and Fugu Ultra, which coordinates a deeper agent pool for accuracy-heavy workflows such as AI research, cybersecurity analysis, and patent investigation. In testing, Fugu Ultra demonstrated frontier-level capability, matching or exceeding models like Gemini 3.1 Pro (high), Opus 4.8 (max), and GPT 5.5 (xhigh) across engineering, scientific, and reasoning benchmarks.

Sakana AI designed Fugu as a safeguard against single-vendor dependency, allowing users to swap agents dynamically if a provider restricts access. The company intends to integrate additional open-source models and future proprietary Sakana AI models into the agent pool to improve coordination and resilience. The system is available now through subscription or pay-as-you-go plans, following a beta program with nearly 500 users. Early feedback indicated high persona stability during long sessions and success in automated research and end-to-end security assessments.

Sakana AI released Sakana Fugu, a new multi-agent orchestration system that functions as a single foundation model. The system dynamically delegates complex, multi-step tasks to a pool of specialized language models, managing selection, communication, and synthesis internally through an OpenAI-compatible API. This architecture builds upon the company's research into learned model coordination, specifically the Trinity and Conductor systems presented at ICLR 2026.

The product is available in two versions: Fugu, designed for high performance with low latency in tasks like coding and interactive services, and Fugu Ultra, which coordinates a deeper agent pool for accuracy-heavy workflows such as AI research, cybersecurity analysis, and patent investigation. In testing, Fugu Ultra demonstrated frontier-level capability, matching or exceeding models like Gemini 3.1 Pro (high), Opus 4.8 (max), and GPT 5.5 (xhigh) across engineering, scientific, and reasoning benchmarks.

Sakana AI designed Fugu as a safeguard against single-vendor dependency, allowing users to swap agents dynamically if a provider restricts access. The company intends to integrate additional open-source models and future proprietary Sakana AI models into the agent pool to improve coordination and resilience. The system is available now through subscription or pay-as-you-go plans, following a beta program with nearly 500 users. Early feedback indicated high persona stability during long sessions and success in automated research and end-to-end security assessments.