Sakana Fugu: When One Model Is Not Enough, Coordinate a Whole Team

Sakana AI just released Fugu, a system that treats a team of AI models as a single model you can call through one API. Instead of picking one provider and hoping it handles every task well, Fugu dynamically assembles a pool of specialized agents, assigns them roles, and coordinates their work. The result: better performance on coding, reasoning, and scientific tasks than any individual model can manage alone.

The pitch is straightforward. You get frontier-level output without locking into a single vendor. You can drop Fugu into tools like Codex for code review or use it for chatbot services. The whole thing sits behind an OpenAI-compatible endpoint, so your integration stays the same even as the underlying models change.

What Fugu actually does

Fugu runs two tiers. The standard Fugu model balances speed and quality for everyday tasks. Fugu Ultra goes deeper, coordinating a larger pool of agents for hard problems like Kaggle competitions, cybersecurity analysis, and paper reproduction. Early benchmarks show Fugu Ultra hitting 73.7% on SWE Bench Pro, which puts it ahead of Opus 4.8 (69.2%) and GPT 5.5 (58.6%).

The system assigns roles on the fly. Thinker agents reason through problems. Workers execute tasks. Verifiers check the output. Fugu learned these coordination patterns through two ICLR 2026 papers called TRINITY and the Conductor, both focused on learned model orchestration rather than hand-designed workflows.

The research behind the coordination

TRINITY uses a lightweight evolved coordinator that orchestrates multiple LLMs over several turns. It assigns Thinker, Worker, or Verifier roles depending on what the task needs. The Conductor takes a different approach. It is trained with reinforcement learning to discover natural-language coordination strategies, designing communication patterns that help diverse LLM pools outperform individual workers on reasoning benchmarks.

What makes this interesting is that nobody programs the coordination rules. The system learns which models work well together and how to route tasks. That is a shift from traditional multi-agent setups where developers manually define who does what.

Why this matters for developers

If you have built with AI APIs, you know the tradeoff. One model is great at coding but weak at reasoning. Another handles complex logic but costs more and runs slower. Fugu's approach lets you use the right model for each subtask without managing multiple API keys, retry logic, or fallback chains yourself.

There is a compliance angle too. Fugu lets you opt specific agents out of the model pool. If a provider does not meet your data privacy requirements, you can exclude it. That matters for teams operating under GDPR or similar regulations, especially since Sakana AI is not yet available in the EU while they work toward compliance.

The catch

Multi-agent coordination sounds great on paper, but it adds complexity. Latency can be higher since multiple models contribute to a single response. Cost is harder to predict because you are paying per token across different providers. And when something goes wrong, debugging a chain of agent decisions is harder than debugging a single model call.

Fugu is also not available in the EU yet. If your team operates in European markets, you will need to wait until Sakana AI completes GDPR compliance work. That is a real limitation for now.

What to watch next

Sakana AI is not the only company chasing multi-agent orchestration. The broader trend is clear: the question is no longer which single model to use, but how to get several models working together. Expect more announcements like this from other providers in the coming months.

If you want to try Fugu, the API is available now through Sakana AI's website. The standard model targets everyday coding and chat workflows. Ultra is aimed at research and high-stakes analysis. Start with the standard tier, test it against your existing setup, and see if the multi-agent coordination actually moves the needle for your use case.