Sakana AI Introduces Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier
Sakana AI has released Fugu Max, which expands the Pareto Efficiency Frontier by orchestrating its largest pool of open and specialized models to date, and Fugu Ultra v2, which pushes peak performance higher than ever before, without the indispensable reliance on the frontier models it orchestrates.
The Two-Dimensional AI Frontier.
The AI industry has largely focused on one metric: building increasingly capable and expensive foundation models. But real-world AI has another critical dimension, cost.
Using a multi-trillion-parameter model for a simple task may deliver capability, but at unnecessary expense. The next generation of AI systems will need to determine not only how to solve a problem, but which model or tool can do it most efficiently.
The focus is therefore shifting toward orchestration that optimizes for both capability and cost.
Two Axes, One Strategy.
Fugu Max and Fugu Ultra v2 share the same orchestration architecture but target different objectives. Fugu Max optimizes for the best possible output at the lowest cost, while Fugu Ultra v2 targets maximum capability on complex, multi-step tasks.
The Fugu Journey.
In a few months, Sakana Fugu has evolved from a beta concept into an enterprise-grade orchestration engine:
April: Beta launch established the multi-agent orchestration thesis.
June: General availability and Fugu Ultra v1 demonstrated performance competitive with closed frontier models on difficult benchmarks.
July: Fugu-Cyber and the Claude Code interface expanded orchestration into cybersecurity and coding workflows.
August: Sakana Chat demonstrated consumer-scale usage, while a partnership with NVIDIA added Nemotron open models.
September: Fugu Max and Fugu Ultra v2 extend the platform across cost and capability.
The progression reflects Sakana’s broader thesis: orchestration can combine specialized agents and models to improve performance, flexibility, and resilience.
Fugu Max: More Models, Lower Cost.
Fugu Max expands Sakana Fugu’s model pool with a broad range of open-weight and specialized models, including NVIDIA’s Nemotron family. Its orchestration engine dynamically routes tasks to the most efficient model capable of solving them.
The result is frontier-level performance at lower cost, with Fugu Max reaching two to six times lower cost than comparable single-model approaches on some workloads.
Performance: Best overall scores across six benchmarks, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish.
Cost: $2 per million input tokens and $6 per million output tokens, with output pricing 40–60% below the cited alternatives.
Efficiency: Extends the cost-performance frontier across seven of ten benchmarks.
Sakana’s broader thesis is that open and specialized models become more powerful when orchestrated together rather than deployed in isolation.
Fugu Ultra v2: Pushing the Capability Frontier.
Fugu Ultra v2 targets maximum capability for complex reasoning, autonomous research, and software development. It is designed to push the performance frontier rather than optimize primarily for cost.
The system performs strongly on tasks involving visual and structured data, scoring 48.3 on Chartography and 74.3 on DeepSWE, while also achieving the best or joint-best score on five of eight benchmarks.
Performance: Best or joint-best on GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon.
Consistency: Top-two performance on seven of eight benchmarks.
Frontier: Pushes the Pareto frontier toward higher capability for quality-sensitive workloads.
Fugu Ultra v2 achieves these results without relying on the cited proprietary frontier models in its agent pool. Instead, it combines open and specialized models, reducing dependence on individual vendors and improving resilience against lock-in, API changes, and service disruptions.
𝐑𝐞𝐦𝐨𝐭𝐞 𝐇𝐢𝐫𝐞 𝐨𝐫 𝐁𝐮𝐢𝐥𝐝 𝐖𝐢𝐭𝐡 𝐔𝐬: 𝐁𝐮𝐢𝐥𝐝𝐰𝐞𝐫𝐤𝐬 helps technology companies build in India: two ways. Hire vetted remote engineers, product leaders, designers, and GTM professionals directly onto your team. Or partner with a proven development studio to design, build, and ship your product end-to-end. Learn More At: menlotimes.com/buildwerks
