What's this about?
Sakana AI upgraded Fugu-Ultra, its "collective intelligence" system that doesn't train one big model but dynamically orchestrates a pool of existing frontier models into ad-hoc teams. v1.1 adds the latest frontier models to the pool and gains 7.9 points; CEO David Ha says it now beats Fable 5 on complex coding and reasoning — without Fable 5 even being in the pool.
What exactly is "model orchestration"?
Instead of asking one genius, Fugu convenes a committee: it learns which models to assemble for a given multi-step task, what roles to give them, and how they should collaborate — including patterns a human wouldn't design. Related to routing (yesterday's Cursor Router picks *one* model per task) but a step further: coordinated teams, one API.
What's the community saying?
709 + 131 likes
- The provocative claim: an orchestrated pool beating a frontier model that isn't in the pool suggests the ensemble adds capability, not just cost savings.
- It's the third routing/ensemble story in three days (Cursor Router, Echo, Fugu) — the layer above the models is suddenly where the action is.
- Skeptics want the benchmark details — "beats Fable 5" claims from vendor charts get discount rates.