Fugu Ultra v1.0 is the first Ultra-tier entry in Sakana AI's Fugu model family, positioned as a quality-first orchestration system rather than a single trained model. Instead of relying on one set of weights, it dynamically routes work across multiple frontier models in an agent pool, aiming to lift performance on difficult programming and reasoning tasks that benefit from collective intelligence. The intended use case is complex, multi-step work such as coding agents, research assistance, and other agentic pipelines where pulling in specialized strengths from different models can produce stronger outcomes than any single backbone on its own.

In practice, Fugu Ultra v1.0 set the baseline that its successor, Fugu Ultra v1.1, was measured against. Sakana reports that the v1.1 orchestration upgrade improved on every benchmark it tracks, with the largest gains on ProgramBench and Terminal Bench 2.1 and a maximum uplift of 7.9 points over v1.0, along with broader improvements in agentic tasks and advanced reasoning, all offered at the same price point as the original release. That trajectory suggests v1.0 is best understood as the foundation of an evolving orchestration stack, well suited to teams that want a coding- and reasoning-focused service and can benefit from a platform whose model pool is refreshed over time without requiring re-integration.