Fugu Ultra is the higher-quality variant in Sakana AI's Fugu family of orchestrator models, designed to prioritize answer quality on the hardest, multi-step problems. Rather than acting as a single monolithic base model, Fugu models are themselves language models trained to interpret user queries and dynamically devise agentic scaffolds that delegate work to a pool of existing frontier LLMs. This design lets the system compose specialized models on a per-task basis, aiming to deliver stronger aggregate performance than any single underlying agent. The Fugu family is exposed through an OpenAI-compatible API, making it straightforward to integrate into existing pipelines without changing client-side assumptions about the interface.
Fugu Ultra's intended use centers on complex workflows such as long-horizon software engineering, code generation, scientific reasoning, and multi-step analytical tasks. On SWE-Bench Pro it is reported at 73.7%, ahead of Claude Opus 4.8 at 69.2%, with Sakana-reported results also claimed on Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, and CharXiv Reasoning. Training combines large-scale fine-tuning with evolutionary algorithms and reinforcement learning to refine the orchestrator's ability to select and coordinate downstream agents. For practitioners, Fugu Ultra is a strong fit when a single high-quality answer matters more than low latency, while the standard Fugu variant is better suited for everyday, latency-sensitive workloads.
