Get the Fugu Playbook

Sakana Fugu looks like one model behind one API — but it's really a small (~7B) controller trained to ROUTE a pool of other companies' models, call them, check them, and even call copies of itself recursively. The engineering is genuinely clever; the "matches Fable 5" headline is genuinely oversold. This field guide is the honest version, beat for beat with the video: what an orchestration model actually is (and how it differs from a single base model AND from a Mixture-of-Experts); how the controller works (route → delegate → verify → synthesize, built on Sakana's Trinity + Conductor research, behind one OpenAI-compatible API); the REAL benchmark table WITH the asterisk — Fugu Ultra genuinely beats Opus 4.8 on SWE-Bench Pro (73.7 vs 69.2), LiveCodeBench (93.2 vs 87.8), GPQA-Diamond (95.5 vs 92.0) and Humanity's Last Exam (50.0 vs 49.8), which it CAN call — but the bigger "matches Fable 5 and Mythos" claim compares against models that are export-controlled and NOT in Fugu's pool, so it's a measured-vs-self-reported comparison, not a head-to-head; the honest caveats (Mixture-of-Agents isn't new, more calls = more cost/latency, and Fugu is only as good as the pool it can reach); a should-you-orchestrate decision chart; and the proven 500-user beta results (14h/H100 autonomous ML research, a 50-week no-look-ahead trading loop from $10k, end-to-end cybersecurity assessments, and 20+ bugs caught in code review vs rivals' ~3). Every number verified against sakana.ai/fugu-release, The Decoder, Nikkei Asia and VentureBeat. Read the asterisk.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.