The Problem: Single-Vendor Fragility
In June 2026, export controls on Anthropic's Fable 5 and Mythos models sent a clear message: relying on a single provider's API for critical infrastructure is a real vulnerability. Access can disappear overnight, leaving enterprises scrambling. This isn't just a theoretical risk—it's a material threat that Sakana AI's new Fugu system is designed to mitigate.
For years, enterprises have built their AI stacks on the assumption that their chosen vendor's API would always be available. But geopolitical tensions, regulatory changes, and even corporate decisions can disrupt access to critical models. When Anthropic's Fable 5 faced export controls, it wasn't just an inconvenience; it was a wake-up call. Companies that had built their entire AI strategy around a single provider found themselves exposed.
Sakana AI's Fugu is a direct response to this vulnerability. Instead of putting all your eggs in one basket, Fugu allows you to tap into a pool of frontier models through a single API. If one model becomes unavailable, Fugu can dynamically route requests to others, ensuring continuity of service. It's not just about having a backup plan; it's about having a resilient, adaptive system that can handle whatever comes its way.
What Fugu Is (and Isn't)
Fugu isn't your typical multi-agent framework with hand-coded routing rules. Instead, it's a foundation model trained to orchestrate other LLMs. Think of it as a single API endpoint that hides complex coordination under the hood. What's more, Fugu can even recursively call itself within its agent pool, adding a layer of flexibility that traditional systems lack.
At its core, Fugu is a foundation model that has been trained to understand the strengths and weaknesses of various LLMs. It doesn't just route requests based on predefined rules; it dynamically orchestrates the models in its pool to achieve the best possible outcome. This means that Fugu can adapt to the specific requirements of each task, ensuring that the right model—or combination of models—is used for the job.
One of the most innovative aspects of Fugu is its ability to recursively call itself. This might sound like a technical detail, but it has profound implications. It means that Fugu can break down complex tasks into smaller subtasks, each of which can be handled by the most appropriate model in the pool. This recursive orchestration allows Fugu to tackle problems that would be beyond the capabilities of any single model, no matter how advanced.
Two Models, One API
Sakana offers two versions of Fugu: the standard Fugu, optimized for balanced performance and low latency, and Fugu Ultra, tuned for maximum answer quality on complex, multi-step problems. Both are accessible through a single OpenAI-compatible API, making it easy to switch between them based on your needs. Whether you're coding, running chatbots, or tackling AI research, Fugu has you covered.
The standard Fugu model is designed for everyday tasks where speed and efficiency are paramount. It's perfect for applications like chatbots, where low latency is crucial, or for coding tasks where quick, accurate responses are needed. Fugu Ultra, on the other hand, is optimized for tasks that require deep reasoning and complex problem-solving. It's the model of choice for research applications, where the quality of the answer is more important than the speed of the response.
Both models are accessible through a single API, which is compatible with the OpenAI API specification. This means that if you're already using OpenAI's models, switching to Fugu is as simple as changing the API endpoint. There's no need to rewrite your code or learn a new API. This compatibility makes it easy to integrate Fugu into your existing workflows, whether you're building chatbots, coding assistants, or research tools.
Benchmark Performance
Fugu Ultra isn't just keeping up—it's matching the performance of Anthropic's Fable 5 and Mythos Preview on rigorous engineering, science, and reasoning benchmarks. It even outperforms heavyweights like Gemini 3.1 Pro, Opus 4.8, and GPT 5.5 on tasks ranging from AutoResearch to Financial Time Series Prediction. And here's the kicker: Fable 5 and Mythos Preview aren't even part of Fugu's agent pool. They're the benchmarks to beat, not the components.
When it comes to performance, Fugu Ultra is in a league of its own. On engineering benchmarks, it matches the performance of Anthropic's Fable 5, one of the most advanced models available. On science and reasoning benchmarks, it goes toe-to-toe with Mythos Preview, another cutting-edge model from Anthropic. But Fugu Ultra doesn't stop there. It also outperforms other heavyweights like Gemini 3.1 Pro, Opus 4.8, and GPT 5.5 on a range of tasks.
What's truly impressive about Fugu Ultra's performance is that it achieves these results without relying on Fable 5 or Mythos Preview as part of its agent pool. These models are used as benchmarks, not as components of the system. This means that Fugu Ultra's performance is a testament to the effectiveness of its orchestration capabilities, not just the quality of the models in its pool.
The Research Lineage
Fugu's innovation doesn't come out of thin air. It's built on two groundbreaking ICLR 2026 papers: TRINITY and Conductor. TRINITY introduces a compact coordinator optimized with evolutionary strategies, while Conductor uses reinforcement learning to design agent-to-agent communication topologies. Together, they form the backbone of Fugu's ability to dynamically orchestrate multiple models.
The TRINITY paper introduces a compact coordinator that is optimized using evolutionary strategies. This coordinator is responsible for managing the interactions between the various models in the pool, ensuring that they work together effectively to solve complex tasks. The use of evolutionary strategies allows the coordinator to adapt and improve over time, learning from its experiences to become more effective.
The Conductor paper, on the other hand, focuses on the design of agent-to-agent communication topologies. It uses reinforcement learning to optimize the way that models communicate with each other, ensuring that information flows efficiently and effectively. This is crucial for the successful orchestration of multiple models, as it allows them to work together seamlessly to achieve the desired outcome.
Together, these two papers form the backbone of Fugu's ability to dynamically orchestrate multiple models. They provide the theoretical foundation and practical techniques that make Fugu's innovative approach possible. Without the insights and advancements presented in these papers, Fugu would not be able to achieve the level of performance and flexibility that it does.
Real-World Beta Results
With around 500 early users in its beta program, Fugu has already proven its mettle. It's surfaced over 20 bugs in code reviews where competitors found only three. It's handled full security assessments end-to-end, from reconnaissance to reporting. And it's even automated data science research with minimal human intervention. These aren't just lab results—they're real-world wins.
The beta program for Fugu has been a resounding success, with around 500 early users putting the system through its paces. The results have been impressive, to say the least. In code reviews, Fugu has surfaced over 20 bugs where competing systems found only three. This is a testament to the system's ability to analyze code thoroughly and identify issues that might otherwise go unnoticed.
Fugu has also proven its worth in the realm of security. It has handled full security assessments end-to-end, from the initial reconnaissance phase to the final reporting stage. This is a complex and demanding task that requires a deep understanding of security principles and the ability to apply them effectively. Fugu's success in this area is a clear indication of its potential to become a valuable tool for security professionals.
Perhaps most impressively, Fugu has demonstrated its ability to automate data science research with minimal human intervention. This is a task that typically requires a high level of expertise and a deep understanding of the subject matter. The fact that Fugu can handle this task with minimal human input is a testament to its advanced capabilities and its potential to revolutionize the field of data science.
Why It Matters for Enterprises
For enterprises, Fugu offers something priceless: AI sovereignty. If a provider restricts access, Fugu dynamically routes around the disruption. No weight merging or shared architectures are needed—models stay independent, and as new frontier models arrive, Fugu can fold them into its pool, passing those gains directly to users. It's not just about performance; it's about resilience.
In today's fast-paced business environment, enterprises need to be able to rely on their AI systems to be available and effective at all times. But as the recent export controls on Anthropic's Fable 5 and Mythos models have shown, this is not always the case. When access to critical models is disrupted, enterprises can find themselves in a difficult position, struggling to maintain their AI capabilities.
Fugu offers a solution to this problem. By providing a single API that can dynamically orchestrate a pool of frontier models, Fugu ensures that enterprises always have access to the AI capabilities they need. If one model becomes unavailable, Fugu can route requests to others, ensuring continuity of service. This dynamic orchestration is not just a backup plan; it's a proactive strategy for maintaining AI sovereignty.
But Fugu's benefits don't stop at resilience. As new frontier models become available, Fugu can fold them into its pool, passing those gains directly to users. This means that enterprises using Fugu can always take advantage of the latest advancements in AI technology, without having to worry about integrating new models into their existing systems. With Fugu, enterprises can future-proof their AI capabilities, ensuring that they always have access to the best possible tools for the job.