Contents
For a decade the AI industry has raced along one axis: make the model bigger. Sakana AI’s latest release argues the axis that actually matters is two-dimensional, capability against cost, and that the winning move isn’t a bigger single model at all. Fugu Max and Fugu Ultra v2, shipped this week, are built around a router that decides which of a pool of smaller, cheaper models should handle a given piece of work, and that can call another instance of itself when a task needs breaking down further. The headline result: Fugu Ultra v2 beats Anthropic’s Opus 5 and Fable 5 on a visual reasoning benchmark while including neither of those models, nor GPT-6 Astra, in its own pool. If that holds up, it says something uncomfortable for labs betting everything on scaling one model: a well-trained conductor can out-perform the frontier soloist it’s conducting around.
One architecture, two goals
Fugu Max and Fugu Ultra v2 aren’t separate products, they’re the same orchestration core tuned for opposite priorities. Fugu Max asks what the best output is at the lowest possible cost: it dynamically routes each request to the leanest model capable of solving it, drawing on Sakana’s largest pool yet of open-weight and specialized models, including the NVIDIA Nemotron family through a partnership announced in August. At $2 per million input tokens and $6 per million output tokens, Sakana says that undercuts Sonnet 5, GPT 5.6 Terra, and Kimi K3 by 40 to 60 percent on output pricing, while still posting the best overall score on six benchmarks including Terminal Bench 2.1 and GPQAD.
Fugu Ultra v2 asks the opposite question: what’s the highest capability reachable on hard, multi-step work, regardless of cost. It’s still the same router, just pointed at a deeper pool and tuned for sustained reasoning over long coding, research, and document tasks. Upgrading from the original Fugu to either variant is a one-line parameter change against the same OpenAI-compatible API, no migration required.
The number that matters: winning without the frontier
The detail worth sitting with is Ultra v2’s benchmark run. On Chartography, a visual reasoning and data-interpretation benchmark, it scores 48.3 against Opus 5’s 27.3 and Fable 5’s 29.5. On DeepSWE, a real-world software engineering benchmark, it scores 74.3, beating models that cost three to five times more per token. Sakana is explicit that Fable 5, Fable 5.1, and GPT-6 Astra are not in Ultra v2’s model pool at all (its training cutoff is August 28). So this isn’t a case of routing to a frontier model and taking credit for its answer. The orchestration layer itself, coordinating open and specialized models plus recursive calls to itself for decomposition, is producing the result.
That recursive step is the part most single-model comparisons miss: when a task is too large for one pass, the router can split it, hand pieces to specialists, call another Fugu instance to decompose a sub-piece further, then synthesize the outputs. It’s less “one model answers” and more “a small organization answers,” which is also why a swappable pool matters commercially, not just technically: if a provider revokes API access, changes pricing, or gets caught in export controls, Sakana can swap the model behind a role without changing the interface above it.
Try it: how would a router split your task
Switch to a long research task with cost optimization off and the simulator leans toward the expensive, recursive route, roughly what Ultra v2 is built for. Flip cost optimization on and it collapses toward the cheapest specialist that can still plausibly do the job, which is the bet Fugu Max is making instead.
What I’d watch next
The interesting question isn’t whether Fugu’s specific numbers hold up under independent benchmarking, it’s whether “orchestration as the product” becomes its own competitive axis alongside raw model scaling. A router that gets smarter about who to call, and that stays useful even when its favorite specialist gets revoked or repriced, is a hedge against exactly the vendor lock-in and geopolitical risk that’s been showing up in other recent stories this month. If a learned conductor really can out-score the frontier models it deliberately excludes from its own pool, expect more labs to start publishing orchestration benchmarks next to their model cards, not just token counts and context windows.