Overview
- Sakana launched two Fugu tiers that expose an OpenAI‑compatible API: Fugu for lower latency and balanced cost, and Fugu Ultra for deeper orchestration focused on answer quality.
- Fugu works as an orchestration layer that selects and routes subtasks to a set of specialist worker models instead of producing every answer from one monolithic model.
- The company published a technical report, code and demos and cited two ICLR 2026 papers as the system's research basis.
- Sakana's benchmarks claim parity or modest gains over top models on engineering, science and reasoning tests, but those results are vendor‑run and need independent replication.
- Engineers and users should expect tradeoffs from orchestration systems, including higher latency and cost, harder debugging and more complex error attribution, plus potential resilience to single‑vendor access losses.