A model whose job is to run other models
Sakana AI has released Sakana Fugu, a product built on an unusual premise: instead of one more giant monolithic model, the headline release is a model whose main skill is orchestrating other models. Fugu presents a full multi-agent system as a single foundation model - you call one API endpoint, and behind it Fugu decides whether to answer directly or assemble and coordinate a team of expert models to handle a complex, multi-step task. Sakana's framing is that the next frontier isn't bigger models but better coordination of collective intelligence: knowing which model to use, delegating planning and execution, and routing around any single model's weaknesses.
How it actually works
The trick is that Fugu is itself a language model, specifically trained to understand when to delegate, how agents should communicate, and how to merge their outputs into one reliable answer - it can even call instances of itself recursively. It handles model selection, delegation, verification, and synthesis internally, so the messiness of a multi-agent setup never reaches your code. The approach builds on Sakana's published research, including two ICLR 2026 papers - Trinity, an evolved LLM coordinator, and Conductor, on orchestrating agents in natural language - and, crucially, the underlying pool of models is swappable rather than fixed.
Two tiers, one API
At launch there are two models, both reachable through a single OpenAI-compatible API. Fugu trades a little quality for low latency and is positioned as the everyday default, dropping into coding and code-review tools like Codex, chatbots, and other interactive services, with the option to exclude specific agents from the pool for data, privacy, or compliance reasons. Fugu Ultra is tuned for maximum answer quality on hard, long-horizon problems, marshalling a deeper bench of expert agents; Sakana says early users leaned on it for AI research, paper reproduction, cybersecurity analysis, and literature and patent investigations.
The benchmark claim, with caveats
Sakana positions Fugu Ultra as shoulder-to-shoulder with Anthropic's Fable 5 and Mythos Preview across rigorous engineering, scientific, and reasoning benchmarks, and says its Fugu models outperform Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5 on a grab-bag of tasks ranging from automated research to mechanical design, one-shot chess, and financial time-series prediction. Two caveats are worth flagging: the comparison numbers are self-reported (with the full set in a technical report on GitHub) and the baseline figures come from the model providers themselves, and Fable 5 and Mythos Preview aren't actually in Fugu's agent pool because they aren't publicly accessible - which is rather the point Sakana is making.
The real pitch: sovereignty
What sets the announcement apart is its explicit political framing. Sakana argues that relying on a single company's API for critical infrastructure, finance, or governance is now a material vulnerability rather than a hypothetical one, pointing directly at the recent export controls that pulled Anthropic's Fable and Mythos models offline. Orchestration, in this telling, is the practical hedge: because Fugu's agent pool is swappable, if one provider restricts access the system dynamically reroutes around the disruption, and the pool can absorb newer and cheaper models - including Sakana's own - over time. The company pitches this as a realistic blueprint for AI sovereignty: frontier capability without betting your stack on access that a single jurisdiction can revoke overnight.
Availability
Sakana Fugu is generally available now, following a beta with close to 500 early users, with subscription tiers for everyday use and pay-as-you-go pricing for heavier and enterprise workloads. Sakana frames this as a starting point rather than a finish line: it plans to expand the pool of expert agents - including open models and its own - and give users more control over how Fugu delegates on their behalf.