Sakana Fugu

Original Article ↗ Hacker News Discussion ↗

What Fugu Is and How It Works

  • Described as an “orchestrator” or “coordinator” LLM that sits in front of multiple models.
  • Regular Fugu appears to route a request to the most suitable model; Fugu Ultra can create multi-step workflows (e.g., one model for math, another for security checks, another for synthesis).
  • It is exposed as a single OpenAI-compatible endpoint, enabling “agent of agents” setups inside existing tools.

Comparisons to Other Multi-Model Systems

  • Frequently compared to OpenRouter Fusion:
    • Fusion = call several models in parallel + a synthesizer step.
    • Fugu = use a dedicated routing model to pick models and sometimes plan sequences.
  • Also compared to Perplexity, Databricks Omnigent, various open-source orchestrators, and simple ensemble methods from classical ML.
  • Some users note similar open-source fusions reportedly match or beat top models at lower cost.

Performance, Quality, and Speed

  • Mixed reports:
    • Some say Fugu-level performance is near frontier models for certain code-review / reasoning tasks.
    • Others find it weaker than leading frontier models for implementation tasks and prone to mistakes.
  • Latency is a recurring complaint; orchestration and multi-step workflows make it feel slow.
  • Technical report is criticized for showing only modest gains over underlying models.

Pricing, Economics, and Alternatives

  • Many find the $200/month tier expensive, especially given reported ~5 hours of usable time and high token burn.
  • Several commenters prefer:
    • Direct access to frontier models at similar or lower effective cost.
    • Very cheap models (e.g., DeepSeek) or local/open-source setups, sometimes orchestrated themselves.
  • Others note subscription fatigue from stacking many paid AI tools.

Research, Architecture, and Future Potential

  • Some see routing/orchestration as the logical next step: combining specialized models and agents can outperform any single model.
  • There’s interest in applying similar ideas to boost smaller, locally hostable models.
  • Concern that frontier labs could integrate equivalent meta-reasoning and make such services obsolete.

Criticism, Risks, and Ethics

  • Skepticism about “frontier-level” marketing and about just replacing one vendor lock-in with another.
  • Frustration over lack of EU availability.
  • Ethical objections to the company’s involvement in defense/military work.