Sakana AI's Fugu Turns a Fleet of AI Models Into One Endpoint
News Summary
Sakana AI, the Tokyo-based AI research laboratory, officially launched Sakana Fugu and Fugu Ultra on June 22, 2026 (Japan Standard Time / Eastern Time UTC-4), delivering a full multi-agent orchestration system accessible through a single, OpenAI-compatible API endpoint — a design the company describes as "one model to command them all."
What Is Sakana Fugu?
Sakana Fugu is not a single model in the traditional sense. It is a multi-agent orchestration system that presents itself to users as one unified foundation model. When a user sends a request to the Fugu endpoint, the system autonomously decides whether to answer directly or to assemble a coordinated team of specialized AI models to tackle the problem. Model selection, delegation, verification, and final synthesis all happen internally and invisibly to the end user.
Fugu is itself a language model — trained specifically to understand how to delegate tasks, manage communication between sub-agents, and synthesize results into coherent, high-quality answers. Notably, Fugu can call instances of itself recursively as part of its orchestration strategy, enabling deeply nested reasoning chains for especially complex problems.
Two Variants: Fugu and Fugu Ultra
At launch, Sakana AI offers two product tiers built on the same orchestration framework.
Fugu is tuned for a balance of strong performance and low latency, making it well-suited for everyday workflows such as coding assistance, code review, and interactive chatbots. Users can configure agent exclusion lists, allowing organizations to satisfy specific privacy or compliance requirements.
Fugu Ultra is optimized for maximum answer quality on difficult, multi-step problems. Early beta users have deployed it for demanding tasks including AI research, academic paper reproduction, cybersecurity analysis, and large-scale patent and literature investigations. According to Sakana AI, Fugu Ultra can drive a full security assessment end-to-end from a single scoped instruction.
Benchmark Performance
Fugu Ultra's published benchmark results are significant. According to Sakana AI's benchmark grids comparing performance across coding, reasoning, scientific, and agentic tasks, Fugu Ultra matches or rivals frontier models across engineering, scientific, and reasoning evaluations — all without requiring users to hold direct access to any single top-tier model provider.
Research Foundations: TRINITY and Conductor
Fugu builds directly on Sakana AI's foundational research into learned model orchestration. The company's TRINITY and Conductor papers — both presented at ICLR 2026 — introduced the core framework for training a language model to dynamically route tasks to the most appropriate expert agents and synthesize their outputs. Fugu is, in effect, the commercial productization of those research breakthroughs, moving the orchestration intelligence from an external controller into the model weights themselves.
Resilience Through a Swappable Agent Pool
One of Fugu's defining architectural properties is its swappable agent pool. Because the system is not tightly coupled to any single AI provider, it can dynamically reroute requests when a given model or vendor becomes unavailable — whether due to policy changes, outages, or other disruptions. This makes Fugu particularly compelling for organizations and developers that require stable, uninterrupted access to advanced AI capabilities without committing to a single upstream dependency. The design reduces the material vulnerability of building critical workflows on a single model provider's API.
Pricing and Availability
Fugu became generally available on June 22, 2026 (Eastern Time UTC-4) and is accessible through console.sakana.ai via a standard OpenAI-compatible API. Three subscription tiers are offered:
- Standard: $20 per month
- Pro: $100 per month
- Max: $200 per month
All tiers include access to both Fugu and Fugu Ultra. Pay-as-you-go plans are available for enterprise and higher-volume workloads. At launch, Fugu is not yet available in the EU/EEA while Sakana AI works toward GDPR and EU-specific regulatory compliance.
Beta Feedback and Early User Reactions
The Fugu beta opened on April 24, 2026 (Eastern Time UTC-4), and approximately 500 beta users contributed feedback before the June 22 general release. User testimonials published by Sakana AI highlighted several strengths: Fugu Ultra's depth and comprehensiveness in code review tasks, strong persona stability across long multi-turn sessions, and the system's ability to autonomously execute end-to-end security assessments from a single high-level instruction.
What's Next
Sakana AI has indicated plans to expand the agent pools available within Fugu and to provide users with greater visibility and control over the orchestration process. As the company grows its roster of integrated models and refines its orchestration engine, Fugu is positioned as an evolving platform rather than a static product. The launch represents both a research milestone and a commercial bet on the idea that orchestration intelligence — rather than any individual model — will define the next generation of AI products.