Now loading...
The artificial intelligence sector has spent a decade focused on developing larger and costlier foundation models, yet the critical advancement lies in optimizing two dimensions: capability and cost. An AI system that employs a multi-trillion-parameter model for simpler tasks is not considered intelligent but inefficient. The future of AI relies on solutions that not only address issues but do so using the most economical approaches available.
Sakana has announced the launch of Fugu Max, which enhances the Pareto Efficiency Frontier by coordinating its most extensive collection of open and specialized models to date. Additionally, the introduction of Fugu Ultra v2 elevates performance benchmarks without an essential dependence on the foundation models it orchestrates.
Fugu Max and Fugu Ultra v2 are variations of the same core orchestration architecture, with each designed for specific missions: Fugu Max aims to deliver optimal outputs at the lowest costs, while Fugu Ultra v2 seeks to achieve peak capability for complex tasks.
Sakana Fugu’s rapid development over the last few months is noteworthy. Initially in beta testing, the platform demonstrated multi-agent orchestration as a unified foundation model, achieved general availability with Fugu Ultra v1, and highlighted its ability to match established frontier models on challenging benchmarks. Subsequent updates showed its effectiveness in real-world scenarios, such as cybersecurity and coding tasks. The platform culminates today with the introduction of Fugu Max and Fugu Ultra v2, reaffirming its core promise of superior performance over isolated models and reliable adaptability through its diverse pool of agents.
Fugu Max broadens the repertoire of models that Sakana Fugu can orchestrate, effectively integrating a high number of open-weight and specialized models, including those from the NVIDIA Nemotron family. By intelligently directing tasks to the most efficient model available, Fugu Max achieves remarkable outcomes while significantly reducing token costs.
Positioned on the Pareto frontier, Fugu Max outperforms elite models at a fraction of the cost, showcasing its capability to deliver superior results on multiple benchmarks while maintaining a pricing structure that’s 40-60% lower than competitors like Sonnet 5 and GPT 5.6 Terra. Moreover, it enhances the cost-performance frontier across seven out of ten benchmarks, showcasing its effectiveness and efficiency amid rapid growth and diversification in the AI ecosystem.
On the other hand, Fugu Ultra v2 emphasizes achieving maximum performance for intricate reasoning tasks, including autonomous research and software development. This model excels in tasks requiring sustained reasoning, particularly on complex data, outperforming competitors significantly on key benchmarks such as SWEFish and Chartography. Fugu Ultra v2 sets a new standard in output quality, attaining top scores in various evaluations.
Significantly, Fugu Ultra v2 does not rely on proprietary models, using a versatile pool of open and specialized options to deliver premier results. This approach mitigates risks associated with vendor lock-in and service interruptions, ensuring users retain control over their processes.
Both Fugu Max and Fugu Ultra v2 are now available via the standard OpenAI-compatible API. Current users can seamlessly upgrade with a simple parameter adjustment, maintaining continuity while accessing new enhancements. Interested parties can explore more by visiting the product page or the console site.
Ultimately, Sakana envisions that the future of AI will not stem from singular models, but rather from collaborative orchestration. With Fugu Max reducing intelligence costs and Fugu Ultra v2 broadening the scope of autonomous execution, Sakana Fugu aims to provide the resilient, vendor-independent infrastructure essential for true AI autonomy.
