Sakana Beats Frontier Benchmarks Without a Frontier Model

The reflex in AI has been to build one enormous model and throw everything at it. Sakana AI, the Tokyo lab founded by former Google researchers, keeps betting the other way. Its new release, Fugu Max, is not a single model at all. It is a learned orchestrator, a system trained to read a task and route it to whichever combination of smaller, open-weight models will handle it best, then stitch the results together.

The commercial hook is the price. According to AI Weekly, Fugu Max lists at $2 per million input tokens and $6 per million output tokens, which Sakana says runs 40 to 60 percent below comparable frontier offerings. The pricing is flat regardless of context length, and the model supports function calling, structured outputs, image and PDF input, and built-in web search. It is available now through OpenRouter.

Routing as the product

What makes Fugu interesting is the architecture rather than any single score. Instead of one set of weights, it draws on a swappable pool of open and specialist models, including Nvidia's Nemotron family, folded in through a collaboration earlier this year. The router can even call instances of itself recursively, breaking a hard task into smaller ones and dispatching each. Sakana's claim is that a smart conductor pointed at a good orchestra can match a soloist that cost far more to train, and do it cheaper.

The approach leans on a healthy supply of capable open-weight models, which is exactly what the past year has produced. DeepSeek keeps pushing long-context open models at low prices, and open coding agents have closed much of the gap with the paid tools. A router only works if the parts it routes to are good, and increasingly they are.

The catch worth naming

Orchestration adds its own complexity. A system that decides which model to call introduces a new place for things to go wrong, and quality now depends on the router's judgment as much as on any component model. Independent testing over time will tell whether Fugu Max holds up across the messy variety of real workloads or shines mainly on the benchmarks it was tuned against.

Still, the direction is worth watching. If the frontier can be approximated by clever routing over open parts, the moat around the largest models gets a little narrower, and the cost of good-enough AI keeps falling. For most businesses, good enough at a fraction of the price is the whole decision.