Sakana launches Fugu Max and Ultra v2: multi-agent orchestration without closed-model lock-in

The boundary that matters isn't model size — it's the Pareto frontier between capability and cost. That is the core argument behind Sakana AI's next-generation orchestration system, Fugu Max and Fugu Ultra v2. Both versions stake out different points on that frontier, and both get there without leaning on a single proprietary frontier model.

Fugu Ultra v2.0 targets multi-step reasoning, autonomous research and full-cycle software development. According to company data, the system achieves the highest score, or ties for it, on five of eight benchmarks: GDP.pdf, Chartography, DeepSWE, Toolathon and an internal benchmark called SWEFish that reflects Sakana's own coding challenges. In seven of the eight tests it places in the top two spots, a signal of broad consistency. The notable twist: Ultra does not rest on one closed model. Instead it orchestrates a swappable pool of open and specialized models. The result, Sakana claims, outperforms closed ecosystems while insulating users from vendor lock-in, API cancellations, geopolitical turbulence and sudden service outages.

The training-data cutoff for Fugu Ultra stands at 28 August 2026. The company explicitly notes that models such as Fable 5, Fable 5.1 and GPT-6-Astra are not included in Ultra's model pool, a detail that delineates the system's current knowledge boundary.

Fugu Max v1.0 expands the pool of models the system can orchestrate, integrating what Sakana describes as an unprecedented number of open-weights and specialized models. A collaboration with Nvidia brought the Nemotron family into the mix. The payoff: Max leads on six benchmarks — Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench and SWEFish — and occupies a point on the Pareto frontier that single-model providers cannot reach. Pricing is set at $2 per million input tokens and $6 per million output tokens.

In practice, Sakana's approach turns orchestration itself into the product. Rather than betting on a single frontier model that may grow more expensive, get blocked or disappear, the system selects the right tool for each step of a task in real time. It is a wager on architectural flexibility over centralized power, and early benchmarks suggest the bet is paying off.