Overview
- Sakana launched Fugu on Monday as a commercial multi-agent orchestration product that developers access through a single OpenAI‑compatible API endpoint.
- Fugu runs a trained 'conductor' model that breaks tasks into substeps, assigns work to specialist agents, verifies outputs, and synthesizes final results in real time.
- Sakana reported that Fugu Ultra scored 73.7 on the SWE‑Bench Pro benchmark and said that result approaches the quality of top standalone models, but that score has not been independently verified in the coverage.
- The service offers two variants — Fugu for low-latency tasks and Fugu Ultra for deeper, complex workflows — and lets developers exclude specific models and opt out of having prompts used for model training.
- Sakana frames Fugu as a practical hedge against export controls and vendor lock-in by routing around restricted providers, a shift that echoes wider industry moves from single, giant models toward coordinated specialist systems and redundancy.