Sakana AI's Fugu Orchestrates Frontier Models

Sakana AI has unveiled Fugu, a new product designed to act as a unified foundation model for multi-agent orchestration. This system dynamically coordinates the world's leading AI models to tackle complex, multi-step tasks, all accessible through a single API. The company claims Fugu Ultra, one of the launch models, rivals the performance of industry benchmarks like Anthropic's Fable 5 and Mythos Preview without the geopolitical risks associated with export controls. The initial announcement highlights Fugu's ability to deliver frontier-level AI capabilities through a collaborative ecosystem rather than relying on monolithic models.

Beyond Bigger Models: The Orchestration Frontier

For years, AI progress has been synonymous with scaling up massive, single models. However, real-world challenges often require a diverse range of specialized knowledge and skills. Sakana AI posits that unlocking peak performance necessitates collective intelligence, knowing which model to deploy, how to delegate tasks, and how to synthesize domain-specific strengths while navigating individual model weaknesses.

This move towards orchestration addresses a growing concern about single-vendor dependency. Recent export control actions have underscored the vulnerability of relying on a single company's APIs for critical functions. Sakana Fugu is engineered to dynamically route around such disruptions by utilizing a pool of swappable agents. This strategy aims to provide a resilient blueprint for AI sovereignty.

What is Sakana Fugu?

Sakana Fugu functions as a multi-agent system that presents itself as a single model. Users interact with one endpoint, and Fugu internally manages model selection, delegation, verification, and synthesis. This means the inherent complexity of a multi-agent setup is abstracted away from the developer's code.

The core innovation lies in Fugu itself being a language model trained to understand delegation, inter-agent communication, and result aggregation. This approach builds on Sakana AI's research in learned model orchestration, including their ICLR 2026 papers on Trinity and The Conductor.

The system offers two models: Fugu, optimized for balance between performance and low latency for everyday tasks, and Fugu Ultra, tuned for maximum accuracy on demanding, multi-step problems. Both are accessible via an OpenAI-compatible API. Early use cases have included AI research, cybersecurity analysis, and patent investigations.

Users are reporting significant advantages in complex workflows. One software engineer noted Fugu Ultra's superiority in code review, surfacing over twenty issues compared to the typical three from other tools. An executive at an enterprise platform company praised Fugu's stable persona maintenance across long sessions, a crucial factor for agent products. A cybersecurity engineer detailed how Fugu autonomously handled a full security assessment from reconnaissance to reporting, staying within scope and avoiding destructive actions.

This demonstrates how Sakana Fugu excels when tasks are messy, long-running, and beyond the scope of a single model call, offering a compelling alternative to traditional multi-agent orchestration system solutions and advancements seen in platforms like multi-agent orchestration system implementations.

Looking Ahead

Sakana Fugu is available starting today. Sakana AI plans to continuously improve the system by incorporating new frontier models and expanding its pool of expert agents. This evolution aims to provide users with increasingly sophisticated and resilient AI capabilities.