Most AI models try to answer every question with their own parameters. Fugu, from Sakana AI, does something different: it does not generate answers. It decides which model should answer.
Fugu is a family of orchestrator models that dynamically coordinates a team of LLMs from different providers - OpenAI, Anthropic, Google, others - for each task. In the benchmarks published by Sakana, Fugu and its Fugu Ultra variant match or surpass frontier models like Claude Fable 5 and Mythos on code, science, and reasoning tests.
The idea has deep implications. If the AI frontier is not only about training larger models, but about combining the existing ones more intelligently, Fugu is a bet on that thesis.
What is Fugu
Fugu is not a traditional generative LLM. It is a multi-agent orchestrator model: for each request, it selects and dynamically combines other models. If one of the internal providers fails, Fugu automatically redirects to another without losing progress on the work.
Sakana offers two variants:
- Fugu: balanced between performance and latency, designed as the default option for daily work.
- Fugu Ultra: optimized for maximum quality on complex multi-step tasks.
Access is via an OpenAI-compatible API, at console.sakana.ai, through OpenRouter, and from the OpenAI Codex CLI.
Benchmark performance
According to tests published by Sakana AI, Fugu and Fugu Ultra approach frontier models on demanding benchmarks:
- LiveCodeBench: Fugu and Fugu Ultra surpass Fable 5.
- SWE-bench Pro: competitive results, with Fugu Ultra matching or improving on Mythos.
- SciCode and GPQA-Diamond (scientific reasoning): Fugu Ultra matches or surpasses Mythos.
These results are notable because Fugu does not generate content on its own: its value comes from better orchestrating other models. It is a layer of intelligence on top of the existing ecosystem, not another model in the list.
The business model
Fugu Ultra’s price ranges from $5 to $10 per million input tokens and from $30 to $45 per million output tokens, depending on context. Fugu, on the other hand, charges the standard rate of the underlying model it ends up using.
That pricing scheme says something about the business model: Fugu charges a margin for orchestration, while the cost of compute is passed through to the chosen model provider. For customers, it turns Fugu into an abstraction layer with a surcharge, but with the promise of a better result-per-dollar ratio than choosing manually.
The academic foundation
Fugu did not emerge from nowhere. The underlying technology comes from two papers accepted at ICLR 2026: TRINITY and Conductor. Both explore routing and coordination between multiple models as an alternative to monolithic scaling.
That academic provenance matters. Sakana AI, founded by former Stability AI researchers, has specialized in alternative architectures: smaller models trained more intelligently, rather than larger models trained with more compute. Fugu is the practical culmination of that philosophy: instead of building another frontier model, build the system that uses them all best.
What it means for the industry
If Fugu delivers what it promises in production, the consequences are several:
- Dependence on a single provider is reduced. An orchestrator that automatically switches models is a form of operational insurance.
- The ROI of frontier training shifts. If frontier models are more accessible via orchestration, training the next largest model stops being the only strategy.
- A new product category appears: the AI orchestrator, analogous to how load balancers transformed web infrastructure.
Fugu is not going to replace frontier models. But it can change who decides which one to use and when.