Sakana AI has launched Fugu, a system it describes as a multi-agent orchestration layer that can be accessed through one OpenAI-compatible model API. The company says the product is designed to make a pool of specialized models appear to developers like a single service, while handling routing and coordination behind the scenes.
In its announcement, Sakana said Fugu is built to manage tasks by selecting, delegating, verifying, and synthesizing work across different language models. The company said the system can answer directly when a request is simple enough, or coordinate multiple expert models when the problem calls for more complex reasoning. According to Sakana, the complexity of the underlying multi-agent setup is hidden from the developer's code.
The launch centers on two offerings. Fugu is positioned as the lower-latency option for day-to-day use, including coding workflows, chatbots, and other interactive applications. Sakana said users can also choose to exclude certain agents from the pool for compliance reasons. Fugu Ultra is the higher-capability version, aimed at difficult multi-step tasks such as AI research, cybersecurity analysis, and patent investigations.
Sakana says both models are available through a single API that is compatible with OpenAI-style integrations. That setup is meant to simplify adoption for developers who already work with common model APIs, while still giving the system flexibility to draw on multiple agents as needed.
The company also framed Fugu as part of a broader argument that orchestration could become as important as building ever-larger standalone models. In its release, Sakana said the concentration of critical AI infrastructure in a handful of providers can create operational and geopolitical risk. It pointed to export controls affecting some frontier models as an example of how access can change quickly.
Under Sakana's description, Fugu is itself a language model trained to call other models in an agent pool, including instances of itself. The company said this allows it to dynamically route tasks to the most suitable model for a given request. Sakana presented the system as a way to combine collective intelligence from multiple models rather than relying on a single monolithic one.
The company compared Fugu's performance to leading frontier systems, saying Fugu Ultra matches the capabilities of models it names Fable and Mythos across engineering, scientific, and reasoning benchmarks. The announcement did not include full benchmark details in the thread, but it said the system is intended to deliver frontier-level capability through orchestration rather than through a single massive model.
Fugu arrives as AI companies increasingly explore techniques that combine multiple models or agents to solve harder problems. Sakana has already been active in that area, recently discussing another approach called AB-MCTS, which used several frontier models together to improve performance on challenging benchmarks. Fugu extends that theme into a productized system meant for direct use by developers.
For now, Sakana is presenting Fugu as a practical interface for multi-agent AI rather than a research demo. The company is pitching the system as a way to gain flexibility, performance, and resilience without forcing users to manage the orchestration logic themselves.