Fugu Ultra v2.0

Fugu Ultra v2.0 is positioned as the capability-first half of a Sakana AI model pair, launched alongside Fugu Max v1.0 and built on a shared orchestration architecture. Rather than being a single monolithic network, the system is designed around learned coordination between sub-models, an approach Sakana has documented in ICLR 2026 work on orchestration. Ultra v2 keeps that same coordination framework but is tuned for complex tasks rather than for the lowest possible inference cost, making it the heavier-duty sibling in the lineup and the natural choice when a workload demands depth over economy.

In Sakana's own reporting, Ultra v2.0 reaches 48.3 on Chartography and 74.3 on DeepSWE, and is described as best or joint-best on GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon, placing it in the top two on seven of the eight benchmarks Sakana highlights. That profile points to a model aimed at multi-step reasoning, software engineering, and tool-driven workflows where orchestration across specialized components matters more than raw single-call fluency. For builders, the practical fit is workloads that benefit from coordinated specialists — long agentic chains, structured multi-stage code tasks, and benchmark-style evaluations — rather than lightweight single-turn interactions.