Definition
Plain language
A system that beats top AI models by acting like a contractor — never solving problems itself, just deciding which model to call for each piece.
As stated in the literature
Sakana's behavioral-level model-merging system that orchestrates closed frontier models via learned routing and workflow composition without touching weights; Fugu picks one worker per turn, Fugu-Ultra composes multi-agent workflows.
Also called: Fugu-Ultra
Why it matters: It shows that smartly orchestrating existing models can outperform any single one, without the cost of retraining or access to their internal weights.
For example, faced with a coding task, Fugu doesn't write the code itself but decides which frontier model is best suited and routes the job to it.
Heard on the show
“The paper is the Sakana Fugu technical report, out June 19th, 2026, and the starting observation is one anyone who uses these models has half-noticed already.”Episode 166 — A Router That Beats the Frontier Models It Calls