About Sakana AI
Founded in Tokyo, Sakana AI represents a fundamental shift in how artificial intelligence can achieve frontier performance. Co-founded by former Google Brain researchers David Ha and Llion Jones—a co-author of the seminal “Attention Is All You Need” paper—our company was built on a simple but powerful insight: the future of AI lies not in ever-larger monolithic models, but in orchestrated collective intelligence.
Our flagship system, Sakana Fugu (officially named Fugu AI), launched in June 2026, embodies this philosophy. Fugu is not a traditional large language model. Instead, it is a “model of models”—an orchestration language model trained to act as a multi-agent coordinator. It dynamically delegates sub-tasks to a curated pool of publicly available frontier models, including GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8, then synthesizes their outputs into a cohesive, high-quality result. All of this is accessible through a single, OpenAI-compatible API endpoint.
Our Mission
We believe that access to world-class AI should not depend on a single vendor or be vulnerable to sudden geopolitical disruptions. Our mission is to democratize frontier AI performance by building systems that are resilient, cost-effective, and transparent. By orchestrating the best available models rather than training one massive, proprietary system, we help organizations mitigate risks of vendor lock-in and export controls—concerns that became urgent following recent U.S. restrictions on advanced AI models.
Our Values
- Collective Intelligence: We believe that the whole is greater than the sum of its parts. By combining the strengths of multiple expert models, we achieve results that rival or surpass even the most restricted frontier systems.
- Resilience and Independence: Our architecture is designed to operate without dependence on any single model provider. Fugu Ultra, for instance, achieves benchmark scores comparable to restricted models like Claude Mythos Preview and Fable 5—without ever using those models in its agent pool.
- Transparency and Trust: We publish technical reports detailing our architecture and benchmark results. We are committed to independent validation and honest communication about both capabilities and limitations.
- Practical Utility: We focus on real-world performance. Whether for everyday coding, AI research, cybersecurity analysis, or patent investigation, our systems are built to deliver tangible value.
Our History
Sakana AI was founded in Tokyo by David Ha and Llion Jones, bringing together deep expertise in neural network research and the transformer architecture that underpins modern AI. The team recognized early that the trend toward ever-larger monolithic models created vulnerabilities: single points of failure, vendor lock-in, and exposure to geopolitical risks.
After years of research into multi-agent coordination and model orchestration, we launched Fugu AI in June 2026 with two primary variants. The standard Fugu balances performance and latency for everyday coding and chat tasks. The flagship Fugu Ultra is optimized for complex, multi-step challenges, orchestrating a deeper pool of expert models to tackle AI research, cybersecurity, and patent investigation.
Our technical reports demonstrate that Fugu Ultra achieves performance comparable to leading restricted models on critical benchmarks—73.7 on SWE-Bench Pro and 95.5 on GPQA-Diamond—entirely without using those restricted models. This validates our core thesis: collective intelligence can match or exceed monolithic approaches.
Today, Sakana Fugu is available as a proprietary API, with pricing at $5 per million input tokens and $30 per million output tokens for Fugu Ultra, along with additional subscription plans. We are currently restricted in the EU/EEA due to GDPR compliance work, and we continue to expand our capabilities while maintaining our commitment to transparency and user trust.
We invite developers, researchers, and enterprises to explore a new path to frontier AI—one built on collaboration, resilience, and collective intelligence.