Sakana AI Releases ‘Fugu Ultra’ to Match Frontier Performance via Autonomous Model Orchestration

Sakana AI, building frontier AI in Japan, led by David Ha, Llion Jones, and Ren Ito, has introduced Sakana Fugu, a new product from Sakana AI that delivers a full multi-agent orchestration system as a single foundation model. 

Fugu dynamically orchestrates the world’s leading AI models to solve complex, multi-step tasks through a single API. This approach enables multi-agent intelligence to deliver frontier-level performance while eliminating single-vendor dependency and avoiding the operational complexity typically associated with traditional multi-agent systems.

Beyond Bigger Models: Orchestration Models Are the Next Frontier.

AI progress has long been driven by larger and more powerful models. However, real-world challenges require diverse expertise that no single model can fully provide. Achieving the best outcomes increasingly depends on collective intelligence: selecting the right models, coordinating specialized tasks, and combining strengths while mitigating weaknesses.

This approach has become both a performance advantage and a strategic necessity. Dependence on a single AI provider exposes organizations and governments to operational and geopolitical risks, as access can be affected by regulatory changes, export controls, or policy decisions.

Sakana Fugu addresses this challenge through orchestration. By dynamically coordinating a network of interchangeable AI models through a single API, Fugu delivers frontier-level performance while reducing single-vendor dependency. As the AI ecosystem evolves, new models can be seamlessly integrated, creating a more resilient and sovereign foundation for advanced AI systems.

What Is Sakana Fugu?

Sakana Fugu is a multi-agent AI system that operates through a single model API. A request is sent to one endpoint, and the system determines whether it can be solved directly or requires coordination among multiple specialized models. Model selection, task delegation, verification, and response synthesis are handled automatically, eliminating the complexity traditionally associated with multi-agent architectures.

At the core of Fugu is an orchestration model trained to determine when to delegate tasks, how agents should collaborate, and how their outputs should be combined into a reliable final answer. Built on Sakana AI’s research in learned model orchestration, including Trinity and Conductor, Fugu delivers the power of a coordinated network of expert models through the simplicity of a single API.

Fugu and Fugu Ultra.

Sakana Fugu is available in two variants, both accessible through a single OpenAI-compatible API.

Fugu is optimized for strong performance with low latency, making it well-suited for coding assistants, code reviews, chatbots, and other interactive applications. Organizations with privacy, compliance, or data governance requirements can also exclude specific agents from the orchestration pool.

Fugu Ultra is designed for maximum performance on complex, multi-step tasks, leveraging a broader network of specialized agents to deliver deeper reasoning and higher answer quality. Common use cases include AI research, paper reproduction, cybersecurity investigations, and patent or literature analysis.

Across industry-standard benchmarks, Fugu Ultra performs competitively with leading frontier models on engineering, scientific, and reasoning tasks, delivering advanced capabilities through an orchestrated multi-model architecture while reducing reliance on any single provider.

What Early Users Are Building?

Benchmark results tell only part of the story. Fugu’s strengths become most apparent in complex, real-world workflows involving multiple steps, uncertainty, and ongoing iteration. Feedback from nearly 500 beta users helped refine the system and highlighted its ability to sustain progress across challenging tasks.

One notable use case emerged in automated data science research, where Fugu Ultra demonstrated the ability to explore ideas, run experiments, analyze outcomes, adjust strategies, and continue advancing with minimal human oversight. These open-ended workflows represent the type of problem orchestration models are designed to handle.

Early adoption has also expanded across code review, cybersecurity assessments, paper reproduction, and literature and patent research. In these environments, value comes not from answering a single prompt, but from coordinating long-running processes that involve reading, implementation, testing, evidence gathering, and synthesis into actionable outputs.

The beta program reinforced a key insight: multi-agent orchestration delivers the greatest advantage when tasks are complex, iterative, and difficult to solve through a single model interaction. Today, both Fugu and Fugu Ultra are generally available through a unified API, offering organizations access to advanced AI capabilities across a wide range of demanding workflows.

Looking Ahead.

The launch of Fugu marks the beginning of a broader vision for AI orchestration. Because the system is built on learned orchestration rather than fixed workflows, its capabilities can continuously improve as the AI ecosystem evolves. New frontier models, open-source models, and future Sakana AI models can be integrated into the agent network, allowing performance gains to be delivered without requiring changes to user workflows.

Future development will focus on expanding the pool of expert agents, improving coordination across long-running and agentic tasks, and providing greater transparency and control over how work is delegated and executed. As orchestration models mature, the goal remains the same: combining the strengths of the world’s best AI systems into a single, resilient, and increasingly capable platform.