Japan's Fugu AI Swarm Just Crashed the Frontier Party
One model is so last year. Japan's Sakana wired a swarm of them together into Fugu... and the swarm is now standing shoulder-to-shoulder with the frontier by making the model pool itself part of the product.
- 73.7SWE-Bench Pro
- 1 APIswarm interface
- 2Fugu models
Sakana's Fugu is a useful reminder that the AI race is not only about building the largest single model. Sometimes the product is the system that knows which model to call, how to coordinate the answer, and how to make the result feel like one intelligence.
Fugu is presented as a multi-agent system available through one OpenAI-compatible API. Behind that single endpoint, Sakana can coordinate multiple models and expert agents instead of asking one monolith to be good at every task.
What happened
- Sakana launched Fugu and Fugu Ultra as multi-agent systems behind one API.
- Fugu Ultra is positioned for harder, higher-stakes problems that benefit from a deeper expert-agent pool.
- Sakana reports a 73.7 score for Fugu Ultra on SWE-Bench Pro.
- The company says Fugu sits near restricted systems such as Fable 5 and Mythos Preview on several benchmarks.
- Fable 5 and Mythos Preview are not part of Fugu's agent pool because they are not publicly accessible.
The SystemOne call, many agents
The exact leaderboard matters less than the product thesis. If the intelligence sits in the orchestration layer, the system can improve as the model pool improves. A single-model product has to wait for its next genius. An orchestrator can route, compare, synthesize, and recruit whatever expert is strongest for the job.
- Routing
- Synthesis
- Specialist agents
- Fallback paths
The ScoreboardBenchmarks with pressure
The benchmark story is what made the announcement travel. The numbers are still vendor-reported, so they deserve pressure before they become mythology. But the direction is hard to ignore: the wrapper is becoming part of the intelligence, not just packaging around it.
- SWE-Bench Pro
- LiveCodeBench
- GPQA-D
- Agentic research tasks
The ThesisThe pool becomes product
Fugu's real story is that the model pool can become the product. The next frontier may not look like one brain getting larger. It may look like many capable systems being conducted well enough that the user never sees the orchestra.
- One interface
- Many model strengths
- Export-control context
- Orchestration as moat