Fugu: Sakana's AI Orchestrator That Routes Around Export Restrictions
Japanese startup Sakana has launched Fugu, a multi-agent orchestration system designed to help developers and enterprises avoid vendor lock-in and navigate export controls on advanced AI models. Named after the Japanese word for 'pufferfish', Fugu dynamically routes user queries to a pool of swappable AI agents, each powered by different large language models (LLMs) from various providers.
The system aims to provide a more reliable alternative to monolithic AI models. If one model provider faces an outage or regulatory restrictions—such as recent US export controls that forced Anthropic to cut off access to its Claude Fable 5 and Mythos 5 models for foreign nationals—Fugu can automatically reroute queries to other LLMs in its pool, minimizing disruption.
Sakana CEO David Ha stated on X: 'We are proving that a well-orchestrated pool of swappable agents can match restricted frontier models like Fable and Mythos. Relying on a single company's model for national infrastructure is a massive risk. As recent export controls have shown, access to top models can disappear overnight. Collective intelligence is the practical hedge against this concentration of power. Fugu simply routes around vendor restrictions by relying on an entirely swappable agent pool.'
Unlike a standalone foundational model, Fugu acts as a multi-agent coordinator. It breaks down user requests into sub-tasks, delegates them to a pool of LLM-powered agents, verifies the work, and synthesizes the final output. Users access Fugu through a standard API endpoint and do not directly interact with the agent swarm.
Sakana has not disclosed the exact number of models in Fugu's pool or the specific routing logic, stating that this information is proprietary and hidden from users by design. However, developers can exclude certain models or providers to comply with corporate privacy standards, and users can opt out of having their prompts used as training data.
On third-party benchmarks, Sakana claimed that Fugu Ultra matched the output quality of frontier models like Claude Fable 5 and Mythos 5. On LiveCodeBench, a coding benchmark, both Fugu (92.9) and Fugu Ultra (93.2) scored higher than Claude Fable 5. The company emphasizes that the system's strength lies in its orchestration rather than any single model.
While Fugu offers a promising approach to circumventing export controls and vendor dependence, its reliance on undisclosed models and proprietary routing raises questions about transparency and reproducibility. As AI export restrictions evolve, such orchestration systems may play a growing role in ensuring continued access to advanced capabilities.
