Japanese AI startup Sakana has unveiled Fugu, a multi-agent orchestration system that achieves frontier-level performance by dynamically routing queries to a pool of specialized language models via a single OpenAI-compatible API. Fugu, meaning 'pufferfish' in Japanese, is designed to provide resilience against vendor lock-in and geopolitical export controls, a pressing concern after Anthropic revoked access to its Claude Mythos 5 and Claude Fable 5 models on June 12 following a U.S. government order.
Fugu surpasses Claude Fable 5 on LiveCodeBench and GPQA-D benchmarks
According to Sakana's benchmarks, Fugu Ultra scored 93.2% on LiveCodeBench, an open-source coding benchmark with regularly refreshed tasks, outperforming Claude Fable 5's 89.8%. The standard Fugu variant scored 92.9%. On GPQA-D, a test of 198 graduate-level multiple-choice questions in biology, physics, and chemistry, both Fugu and Fugu Ultra achieved 95.5%, edging out the prior Claude Mythos Preview's 94.6%. These results indicate that a well-orchestrated swarm of models can match or exceed the capabilities of the most advanced monolithic models, especially on complex, multi-step tasks.
Sponsored Protocol
How Fugu orchestration works and its enterprise value
Unlike simple model routers that dispatch a single request to the best-fit model, Fugu operates as a general contractor: it decomposes a task into subtasks, delegates them to specialized models, verifies outputs, and synthesizes a final result. The system is grounded in Sakana's 2026 research papers TRINITY and Conductor, and autonomously manages coordination using learned strategies. CEO and co-founder David Ha, formerly of Google Brain, stated that 'orchestration models are the next frontier, beyond bigger models.' In an era where access to top AI models can vanish overnight due to regulations or export controls, Fugu offers a practical hedge: it can route around restrictions by swapping its agent pool.
Sponsored Protocol
Pricing, availability, and the EU exclusion
Fugu is available immediately in most regions, with a temporary exclusion of the European Union and European Economic Area while Sakana aligns its black-box routing architecture with GDPR. Pricing includes monthly subscriptions (Standard $20, Pro $100, Max $200) and a pay-as-you-go plan. Fugu Ultra costs $5 per million input tokens and $30 per output, placing it among the higher-priced options, comparable to OpenAI's GPT-5.5. For context windows above 272K tokens, rates rise to $10 and $45 respectively. Notably, background tokens consumed by orchestration are also billed, a factor enterprises must consider when modeling costs.
Strategic implications for resilient AI infrastructure
Sakana's approach marks a shift from monolithic models to collective intelligence, reducing dependency on single vendors. This is particularly valuable for companies and nations seeking to mitigate geopolitical and regulatory risks. As discussed in our article on Meta AI Workers Revolt, managing tech resources and talent is key to innovation. Fugu enables building AI stacks with built-in redundancy: if one provider is compromised, the system adapts. However, the proprietary routing logic and lack of transparency about which models are used raise concerns about data sovereignty, as noted by some developers.
Sponsored Protocol
For more on multi-agent systems, see Wikipedia.
