Sakana AI Fugu Ultra routes frontier models to rival Fable and Mythos

Fugu Ultra is a shot across the bow of the “one frontier model to rule them all” mindset. Rather than burning billions on training yet another maximalist model, Sakana is betting that a relatively small, reinforcement‑learned router can squeeze more effective capability out of the existing frontier zoo. If their published numbers hold up, a 7B‑parameter conductor that knows when to call which model could deliver Opus‑ or Mythos‑class performance without ever owning a single hyperscale GPU cluster.

Strategically, that matters because it attacks the concentration of power in a handful of US and Chinese labs from the side. A well‑designed orchestration layer can, in principle, arbitrage between OpenAI, Anthropic, Google, local sovereign models and open weights, routing around export controls and vendor lock‑in. Japan, which has struggled to field its own frontier‑scale model, may find this “collective intelligence” approach a more realistic way to stay relevant.

For the AGI race, Fugu doesn’t create a new mind so much as a kind of meta‑cortex over existing ones. But if orchestration systems like this become standard, they effectively raise the system‑level capability of today’s models—especially on complex, multi‑step tasks—without waiting for the next big training run. That tilts the curve toward faster practical progress even if raw model IQ advances more slowly.