What is Fugu, Sakana’s new AI-agent orchestrator to help navigate export controls?

Export controls and the push for AI sovereignty are creating new opportunities for startups looking to challenge the AI industry’s dominant players. Earlier this week, Japanese startup Sakana launched a new multi-agent orchestration system called Fugu that has been designed for developers and enterprises to strengthen their AI capabilities against vendor lock-ins and export restrictions.

Named after the Japanese word for ‘pufferfish’, Fugu takes a different approach from the traditional monolithic model structure. It is an AI system that dynamically routes user queries to a swappable pool of specialised AI agents, according to Sakana.

Fugu has been billed as a more reliable option for enterprise workflows since these agents are not powered by any single large language model (LLM) and, instead, orchestrates multiple LLMs from different providers. This ensures that if one model provider faces an outage or faces sudden regulatory restrictions, Fugu will simply route queries to different LLMs in order to avoid disruption.

This approach could prove particularly relevant in light of US export controls on advanced AI systems. Earlier this month, Anthropic’s Claude Fable 5 and Mythos 5 models became subject to new restrictions, with the AI startup forced to cut off access to the models for foreign nationals following a US Commerce Department directive citing national security concerns.

“We are proving that a well-orchestrated pool of swappable agents can match restricted frontier models like Fable and Mythos,” David Ha, the CEO and co-founder of Sakana, wrote in a post on X. “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,” Ha further said.

Under the hood of Fugu

In contrast to a standalone foundational AI model, Fugu is designed to act as a multi-agent coordinator that relies on a “diverse pool of powerful models”, according to Sakana. “Fugu is itself an LLM, trained to call various LLMs in an agent pool, including instances of itself recursively,” the company said in its technical release.

When a user makes a request, Fugu does not attempt to complete the task on its own. Instead, it breaks down the problem into sub-tasks, delegates them to a pool of LLM-powered AI agents, verifies the work, and synthesises the final output. In this way, Fugu autonomously manages the entire process of selecting models and verifying their work through learned coordination strategies.

Story continues below this ad

Users do not directly interact with the multi-agent swarm, instead accessing it through a standard Application Programming Interface (API) endpoint.

While Sakana has said that Fugu relies on multiple LLMs or “specialised models”, the company has not disclosed how many models are involved. The specific models Fugu selects and how it coordinates them are also unclear, with Sakana stating that the ‘proprietary’ routing information is hidden from users by design.

However, developers have the option to exclude specific models or providers from their routing pool in order to ensure compliance with corporate privacy standards. Additionally, users can also opt-out from having their prompts used as training data for future AI systems.

Benchmark performance

When evaluated on third-party benchmarks of agentic tasks, Sakana claimed that Fugu Ultra matched the output quality of frontier AI models, including Claude Fable 5 and Mythos 5.

Story continues below this ad

Benchmark chart of Fugu and Fugu Ultra. (Image: Sakana AI)

On LiveCodeBench, an open-source benchmark testing coding performance, both Fugu (92.9) and Fugu Ultra (93.2) achieved scores higher than Claude Fable 5. Both models (95.5) also beat the prior Claude Mythos Preview model (94.6) on GPQA-D (Diamond), a test of 198 graduate-level multiple-choice questions in biology, physics, and chemistry.

Pricing and availability

Sakana has said that users can choose between two variants of Fugu. While Fugu is a high-speed, low-latency default model optimised for everyday tasks, Fugu Ultra is capable of carrying out complex, high-stakes tasks such as AI research, cybersecurity analysis, and multi-step patent investigations by coordinating a deeper pool of AI agents.

Notably, Fugu is not an open AI system. It is offered as a commercial, proprietary API service.

The standard Fugu model is accessible via a pay-as-you-go plan with dynamic rates based on the specific underlying models that are activated. Fugu Ultra, on the other hand, comes with a fixed pricing structure starting at $5 per million input tokens and $30 per million output tokens.

Story continues below this ad

Both models are available in most regions except within the European Union (EU) and European Economic Area (EEA) as the company is working to ensure that its black-box data routing architecture complies with the bloc’s General Data Protection Regulation (GDPR).