Sakana AI just dropped what might be the most interesting architectural bet in the AI race right now. The Tokyo-based research lab unveiled Sakana Fugu, a multi-agent orchestration system that doesn’t try to be the biggest model in the room. Instead, it conducts a symphony of specialist models, including third-party frontier LLMs, through a single OpenAI-compatible API.
The headline number: Fugu Ultra, the system’s top-tier variant, scored 73.7 on the SWE-Bench Pro benchmark. That reportedly matches or approaches the performance of Anthropic’s Fable 5 and Mythos Preview. In English: a system that coordinates multiple AI agents is now performing on par with some of the most powerful standalone models on the planet.
How Fugu actually works
Fugu operates as a trained conductor model that manages an entire orchestra of specialist AI agents. Unlike traditional multi-agent setups that rely on fixed, hard-coded workflows, Fugu dynamically handles routing, role assignments, verification steps, and result synthesis. The system learns collaboration patterns autonomously, deciding in real-time which specialist model should handle which part of a complex task.
The whole thing is delivered through a single API endpoint that’s compatible with OpenAI’s format. For developers, that means integration doesn’t require rearchitecting existing applications. You swap out an endpoint and suddenly have access to a coordinated multi-agent system instead of a single model.
