Japanese startup Sakana AI has launched Sakana Fugu, a new kind of AI system that doesn't try to be one giant model. Instead, according to MarkTechPost, Fugu is an "orchestration model" that routes each task to the best tool in a swappable pool of frontier large language models.
The idea is closer to a smart dispatcher than a single brain. MarkTechPost reports that Fugu and a more powerful version, Fugu Ultra, send coding, reasoning, and agentic tasks to different underlying models—and that this approach leads on most of those benchmarks.
The claims are notable because of who Fugu is reportedly beating. NDTV reports that the Japanese system beat Claude 5 on certain benchmarks. Nikkei Asia reports that Sakana Fugu, described as a multiagent AI, scores well against Fable 5 and GPT 5.5—putting it in the same conversation as the top models from the biggest AI labs.
Because the model pool is "swappable," the system isn't locked to any single provider. As stronger models arrive, an orchestration layer can in principle plug them in, rather than requiring a costly retrain from scratch.
A caveat worth keeping in mind: these results are benchmark scores reported by the company and relayed through news outlets, and NDTV's framing is explicitly "reportedly." Independent, real-world testing tends to tell a fuller story than leaderboard numbers.
Why it matters: if a routing-and-orchestration approach can match or beat the frontier models it draws on, the race for the best AI may shift from building one ever-larger model toward smartly combining the ones that already exist.