模型與代理系統
Sakana Fugu Splits Cost and Capability Tracks, Delivers Multi-Model Orchestration Through a Single Compatible API
Sakana AI has launched Fugu Max and Fugu Ultra v2, optimizing multi-model orchestration along two separate tracks: lower cost and greater capability. Both retain an OpenAI-compatible interface, but the official benchmarks still combine Sakana’s own tests with vendor-reported results, providing insufficient evidence of cost advantages in real-world workloads.
Sakana AI released Fugu Max and Fugu Ultra v2 on September 11. Rather than two new sets of open weights, they are two service configurations built on the same learned orchestration architecture: users submit a single request, and Fugu’s conductor model internally selects, delegates to, and coordinates multiple specialized models before synthesizing their outputs. Externally, it still appears as a single model endpoint.
Fugu Max targets the cost–capability Pareto frontier and expands the pool of available open-weight and specialized models, including NVIDIA Nemotron. It is priced at $2 per million input tokens and $6 per million output tokens, with cached input priced at $0.25. Sakana claims that it advances the existing cost–performance frontier on seven of ten evaluations and achieves the highest aggregate score on six benchmarks, including Terminal Bench 2.1, GPQAD, and AutomationBench. Ultra v2 increases orchestration depth for long-horizon reasoning, software engineering, and structured visual tasks. It is priced at $5 per million input tokens and $30 per million output tokens, rising to $10 and $45, respectively, once the context exceeds 272K tokens.
More notably, Ultra v2 scored 48.3 on Chartography, compared with 27.3 for Opus 5 and 29.5 for Fable 5 in Sakana’s comparison; it achieved or tied for the best result on five of eight benchmarks. Sakana also states that its training cutoff was August 28 and that its model pool does not include Fable 5, Fable 5.1, or GPT-6 Astra, meaning it is not simply using the latest closed models as subagents.
The release productizes an alternative scaling path to simply replacing a foundation model with a larger one. Engineering teams can retain the same API while switching between cost-first and capability-first configurations according to workload, and can more readily replace underlying providers. However, most evaluations were conducted by Sakana itself, while SWEFish is an internal benchmark. Cross-model collaboration also inherently increases latency, failure points, and output non-reproducibility. Adopters should therefore evaluate total task cost, tail latency, tool side effects, and retry rates rather than comparing only publicly listed per-token prices.