TOKYO: Tokyo-based AI startup Sakana AI has introduced two new products, Fugu and Fugu Ultra, marking the launch of an entirely new category in the AI space. Rather than competing as standalone foundation models, both products are built around the idea of coordinating multiple frontier AI models through a single API endpoint.
Sakana AI positions Fugu as a product where orchestration itself is the core offering. Instead of depending on one large language model to handle everything, Fugu works as a coordinator that selects, manages and delegates tasks across a pool of specialized AI models, including recursive instances of itself. All of this happens behind the scenes while users interact with a single OpenAI-compatible endpoint.
In simple terms, Fugu decides which AI model is best suited for each part of a task, verifies the outputs and combines them into a single response, without the user needing to manage any of that complexity themselves.
The system builds on two Sakana research papers presented at ICLR 2026 earlier this year. TRINITY introduces an evolved coordinator that assigns Thinker, Worker, and Verifier roles across a model pool, while The Conductor uses reinforcement learning to discover natural-language coordination strategies between agents.
The approach addresses a persistent challenge in enterprise AI development. Most organizations manually design agent workflows, select models for individual tasks, and maintain increasingly complex routing logic. Fugu instead learns how to assemble, coordinate, and optimize a team of AI models dynamically for each request.
Sakana ships the platform in two variants behind the same unified API. Fugu balances performance and latency for everyday coding, code review, content generation, and chatbot applications. Fugu Ultra coordinates a deeper agent pool and additional verification layers for more complex reasoning, scientific analysis, software engineering, and multi-step problem-solving tasks.
According to Sakana AI, Fugu Ultra achieves performance levels comparable to leading frontier systems, including Anthropic’s Fable 5 and Mythos Preview, across several reasoning and engineering benchmarks. The company argues that the orchestration layer itself can outperform the individual models it coordinates by combining their strengths while mitigating their weaknesses.
The lighter Fugu model allows organizations to selectively exclude agents from its pool. That flexibility helps enterprises satisfy data residency, privacy, security, and compliance requirements across different deployments. Fugu Ultra instead operates with a fixed, optimized pool designed to maximize answer quality and reasoning depth.
Sakana frames the architecture as a strategic hedge against single-vendor dependency. Because Fugu sits above the underlying models, enterprises can continue operating even if individual providers change pricing, restrict access, or experience service disruptions. As newer frontier models emerge, they can be incorporated into the orchestration layer without requiring customers to redesign their applications.
The launch also reflects a broader shift within the AI industry. Rather than focusing exclusively on training ever-larger foundation models, companies are increasingly investing in systems that intelligently combine multiple models, tools, and agents into a coordinated workflow. In this vision, competitive advantage comes not only from model capability but from the ability to orchestrate the right model for the right task at the right time.
Sakana is offering a free second month for users who subscribe to any paid tier by July 31. The company also provides an elastic pay-as-you-go plan aimed at enterprise production deployments with higher priority access. Fugu and Fugu Ultra are not available in the European Union or European Economic Area at launch.
Previous
