Fugu Ultra is the high-quality tier of Sakana AI's Fugu family, an orchestrator model designed to understand a user query and then dynamically design an agentic scaffold that calls on a team of external frontier LLMs. Rather than acting as a single self-contained foundation model, it functions as a learned coordinator that composes the specializations of many underlying models into a collectively intelligent system, with the goal of pushing performance past what any individual LLM can reach on its own.

According to the Sakana Fugu Technical Report, Fugu Ultra is targeted at the hardest problems where answer quality matters more than latency, while a sibling Fugu model is tuned for everyday use. The family is trained through a combination of large-scale fine-tuning, evolutionary algorithms, and reinforcement learning, and is reported to achieve state-of-the-art results among publicly accessible models on benchmarks such as SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, and CharXiv Reasoning. It is delivered through the Sakana console and distributed via third-party gateways, making it a practical fit for workflows that need frontier-level reasoning and coding assistance without committing to a single upstream provider.