Today, Sakana AI introduced Sakana Fugu, a multi-agent orchestration platform that operates as a single unified model. You submit a request to one endpoint, and Fugu handles everything behind the scenes. For simpler tasks, it responds directly. For more complex challenges, it automatically assembles and manages a team of specialized models. The intricacies of this multi-agent architecture remain completely invisible to your application code.
Key Takeaways
- Fugu provides a full multi-agent system accessible through a single OpenAI-compatible API.
- Fugu Ultra currently leads on most published coding and reasoning benchmarks.
- The orchestrator consistently outperforms the individual models it manages.
- Opt-out controls and provider routing address compliance concerns and reduce reliance on any single vendor.
- Routing logic is proprietary, meaning the specific model chosen for each query remains hidden.
What is Sakana Fugu?
At its core, Fugu is a language model trained to invoke other LLMs from an agent pool—including recursive calls to instances of itself. Internally, it handles model selection, task delegation, result verification, and final synthesis.
Rather than relying on rigid, pre-defined roles or workflows, Fugu learns how to coordinate dynamically. It determines when to delegate tasks and how agents should collaborate, then merges their outputs into a single coherent response. From your perspective, you’re simply calling one model. Under the hood, a coordinated team of experts does the heavy lifting.
Sakana AI positions this approach as a safeguard against dependency on a single provider. If one vendor restricts access, Fugu seamlessly routes around the disruption. The team points to recent export controls affecting Anthropic’s Fable and Mythos models as key motivation. Over time, newer models can be integrated into the pool.
Fugu and Fugu Ultra: Two Models, One API
Fugu is available in two versions, both accessible via the same OpenAI-compatible API:
- Fugu balances strong performance with low latency, making it ideal for everyday coding, code review, and chatbot applications. It also integrates well with tools like Codex. You can exclude specific agents from its pool, helping teams meet data privacy and compliance requirements.
- Fugu Ultra is optimized for maximum accuracy on complex, multi-step problems. It coordinates a deeper pool of expert agents. Its agent pool is fixed, so opt-out functionality is not available. The current model ID is
fugu-ultra-20260615.
The Research Behind the Orchestrator
Fugu builds upon two ICLR 2026 papers—Trinity and Conductor—focused on learned orchestration.
TRINITY employs a lightweight, evolved coordinator that operates across multiple turns, dynamically assigning Thinker, Worker, or Verifier roles. Conductor is trained using reinforcement learning to discover natural-language coordination strategies and targeted prompts for diverse LLM pools.
Together, these approaches demonstrate that systems can learn to assemble and route agents on a per-task basis, replacing manually designed workflows.
Interactive Explainer
