Sakana AI Launches Fugu: An Orchestration Model Matching Frontier Performance Without Export Risk

The Japanese company Sakana AI, based in Tokyo, announced on Monday, June 22, 2026, the launch of its new system Sakana Fugu, which adopts an approach that diverges from the industry's race to build ever-larger models. Fugu (meaning "pufferfish" in Japanese) is not a single giant model, but a multi-agent orchestration system that behaves and is used like a single model, dynamically distributing tasks across a swappable pool of the world's best models, through a single OpenAI-compatible API.

How Does Fugu Work?

The core idea is that Fugu itself is a language model specifically trained to call and coordinate other models, not just a simple router. You send your request to a single endpoint, and Fugu decides internally how to handle it: solving it directly if that is enough, or assembling a team of expert models and coordinating among them when the task requires it. Fugu internally handles model selection, task delegation, result verification, and synthesis, so the complexity of the multi-agent system never reaches the developer's code. More intriguingly, it can call instances of itself recursively, reading its previous output and launching a corrective workflow.

Two Variants for Different Needs

Fugu comes in two versions through the same API. The first is standard Fugu, aimed at everyday tasks like code review and chat, balancing strong performance with low latency, and allowing the developer to opt specific models or providers out of the routing pool for privacy and compliance purposes. The second is Fugu Ultra (with the ID fugu-ultra-20260615), tuned for maximum quality on complex multi-step problems, coordinating a deeper pool of agents, but its pool is fixed so opt-out is not available.

The Motivation: Hedging Against Single-Vendor Dependency

Notably, Sakana explicitly frames Fugu as a "hedge" against the risks of relying on a single vendor, citing directly the export controls the US government imposed on Anthropic's Fable 5 and Mythos models on June 12. David Ha, CEO and co-founder (formerly of Google Brain), says that relying on a single company's model for national infrastructure is a "massive risk," and that access to the best models "can disappear overnight" due to regulatory restrictions. Accordingly, if a provider restricts access, Fugu automatically reroutes tasks around the disruption through a fully swappable agent pool, thereby building native redundancy into the AI layer.

Performance Claims... With an Important Caveat

Sakana says Fugu Ultra stands "shoulder-to-shoulder" with leading models like Fable 5 and Mythos Preview across the hardest engineering, science, and reasoning benchmarks, and that the standard version outperforms publicly available models on several benchmarks. But honesty requires a fundamental caveat: these comparisons rely on Sakana's own numbers, and the competing models' scores are taken from self-reported data from their developers, not verified by a neutral, independent party. Indeed, Fable 5 and Mythos are not even in Fugu's pool, because they are no longer publicly available after the export controls. The routing mechanism itself is also entirely proprietary and hidden, so the user does not know which model actually handled their request.

Pricing and Availability

Both models are available to enterprise clients today. Fugu Ultra adopts fixed pricing per million tokens: 5 dollars for input, 30 for output, and 0.5 dollars for cached input, rising to 10, 45, and 1 dollar respectively for extreme workloads exceeding a 272K-token context window, placing it among the relatively more expensive options. The standard version is charged at the rate of the highest model used in the request without additional stacked "agent fees." The company ran a beta program of about 500 users before the public launch, and one engineer reported that Fugu Ultra surfaced more than twenty issues in a code review where others surface about three.

What Does This Mean for the Industry?

Fugu signals a possible shift in AI philosophy: from "the model is the product" to "orchestration is the product." It builds on two Sakana research papers at the ICLR 2026 conference (Trinity and Conductor) on learned model orchestration. As models become swappable components, developers may care less about which model temporarily leads a leaderboard, and more about the system's overall ability to route intelligently and remain resilient. But observers also warn that a hidden orchestration layer may obscure the real cost, the source of the answer, and the failure modes, making transparency and observability an essential condition for adopting it in sensitive environments.

Conclusion

The launch of Sakana Fugu is a telling moment for the maturing of "orchestration models" as a product category in their own right, and for the shift of AI supply-chain resilience from a luxury to a strategic necessity, especially for nations and institutions worried about export controls. Independent verification of the performance claims, and transparency about what happens inside the "black box," remain the real test that will determine whether Fugu is a fundamental shift or merely an elegant wrapper for model routing.