Sakana AI has unveiled Fugu, a system that manages other artificial intelligence models rather than competing as a standalone model itself, as the United States and China vie for dominance in the field.
Sakana AI was founded by former Google researchers, and one of its co-founders helped write the original research paper underpinning modern generative artificial intelligence.
The Tokyo-based lab has built its reputation on the idea that a coordinated team of smaller, specialised models can outperform a single giant one.
How Fugu works
Rather than generating answers itself, Fugu acts as a "boss," choosing the most suitable model for each part of a task, running several models at once, and merging their outputs into a single response.
Users interact with it through one connection, much as they would with any standalone chatbot, without seeing the coordination happening underneath.
How it performs
Sakana's own testing found Fugu Ultra scored 73% on SWE-Bench Pro, a demanding coding benchmark, against 58% for OpenAI's GPT-5.5.
Those figures come solely from Sakana's internal testing and have not been independently verified by outside researchers.
Mixed reactions in practice
Some users have reported that Fugu can be slow, with certain tasks taking up to half an hour to complete, a trade-off attributed to the extra time needed to coordinate multiple models before returning an answer.
A Wharton professor of artificial intelligence voiced scepticism about the system, saying it did not feel as powerful in practice as the models it claims to beat on paper.
An open question
The launch lands as US and Chinese firms continue to compete for leadership in frontier AI development, with Sakana positioning its orchestration approach as a way to reach competitive results without building ever-larger models from scratch.
Whether that represents a genuine shift in how AI systems are built, or simply a technique that fails to hold up over time, remains an open question.
