Orchestration de modèles apprise
A model that does not do the work itself, it drives a team of other expert models behind a single endpoint. How tasks get split (who answers, in what order, how to combine the answers) is not hand-coded with if/else rules, it is learned during training. Sakana AI popularized the idea on June 22, 2026 with Sakana Fugu, an orchestration model built on its Trinity and Conductor research (ICLR 2026). Fugu Ultra beats frontier models on SWE-Bench Pro (73.7% versus 69.2% for Claude Opus 4.8) by orchestrating other models instead of getting bigger. The right picture: you move from one very strong expert to a team of experts led by a conductor, who picks the specialists, splits the work, checks the answers and writes the synthesis, behind an API that presents itself as a single model. Sakana also benchmarks Fugu Ultra against Fable 5 and Mythos Preview from Anthropic on coding, reasoning, science and agent tests, noting that neither is part of the Fugu team, since neither is publicly accessible.
Strengths
- Moves performance from model size to coordination quality
- Orchestration is learned, not a cascade of if/else rules to maintain by hand
- A single OpenAI-compatible endpoint, the model pool stays swappable behind it
Limitations
- Hidden routing creates an opaque dependency: you no longer know which model handled your data
- Brand-new frontier product, production hindsight is still missing (latency, real cost, long-run reliability)
Best for
- A tech lead already orchestrating agent teams who wants to see where coordination is heading
- A PM who has to weigh an AI budget without believing only the biggest model matters