Overview
- Sakana publicly launched Fugu and published a technical report on arXiv on Wednesday, June 24, 2026, describing the model training, orchestration methods, and infrastructure behind the product.
- Fugu is itself an LLM trained to act as an orchestrator that breaks requests into subtasks, selects specialist agent models for each subtask, verifies results, and synthesizes a final answer through a single OpenAI‑compatible API.
- Sakana says its high-end tier, Fugu Ultra, scored 73.7 on the SWE‑Bench Pro coding benchmark and reports state‑of‑the‑art results across multiple tests by combining agent strengths rather than relying on one monolithic model.
- The API is commercially available worldwide except in the EU and EEA while Sakana works on GDPR compliance, and the company publishes two tiers and pricing with Fugu Ultra listed at $5 per million input tokens and $30 per million output tokens; routing logic and precise agent composition remain proprietary but developers can exclude providers and opt out of training‑data use.
- Early user reports are mixed, with some praising task-specific gains and others noting slower performance, high token burn and cost concerns, and analysts say Fugu can lower vendor‑lock‑in risk but still depends on external model providers and opaque routing, building on Sakana’s prior TRINITY/Conductor research and recent export‑control debates.