Fugu Ultra is the high-quality tier of Sakana AI's Fugu orchestrator family, designed to coordinate multiple expert LLM agents rather than serve as a single standalone model. According to the Sakana Fugu technical report, these orchestrator models are themselves language models trained to understand user queries and dynamically devise agentic scaffolds, combining fine-tuning, evolutionary algorithms, and reinforcement learning to amplify the combined capabilities of an LLM agent team. While the base Fugu model balances performance with latency for everyday use, Fugu Ultra prioritizes answer quality on the hardest problems by routing between one to three agents depending on the difficulty of the task.
On benchmark leaderboards, Fugu Ultra achieves state-of-the-art results among publicly accessible models across challenging evaluations including SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, and CharXiv Reasoning. Independent reporting from The Decoder notes that the system can match leading proprietary benchmarks, though hands-on testers observed significant latency on complex tasks and pointed out that the pool still depends on external frontier providers. Fugu Ultra fits workflows that demand depth over speed, such as multi-step coding agents, complex reasoning pipelines, and high-stakes analytical tasks where the goal is the strongest available answer rather than the fastest response.
