Sakana Trained One AI to Command GPT-5.5, Opus, and Gemini — It Cracked 73.7 Where They Stalled at 69
Last Updated on June 25, 2026 by Editorial Team
Author(s): Chew Loong Nian – AI ENGINEER
Originally published on Towards AI.
Sakana Trained One AI to Command GPT-5.5, Opus, and Gemini — It Cracked 73.7 Where They Stalled at 69
Two days ago a Tokyo lab shipped a model that scored 73.7 on SWE-Bench Pro. Opus 4.8 gets 69.2 on the same test. GPT-5.5 gets 58.6. Gemini 3.1 Pro gets 54.2. Here is the part that broke my brain: the model that scored 73.7 has no weights of its own for that task. It is GPT-5.5, Opus, and Gemini — the exact models it just beat — wearing a trench coat.
After the lead, the article explains that Sakana’s “Sakana Fugu” is an LLM trained to coordinate other LLMs—delegating subtasks to different models, running verification passes, and stitching results together inside a single endpoint—then walks through the benchmark claims (including why orchestration wins on many rows but not all). It argues the real story is the export-control disruption that cut off access to specific frontier models, positioning Fugu’s swappable “agent pool” as resilience against sudden provider loss. The piece also covers criticisms that Fugu may be “OpenRouter with extra steps,” that benchmark comparisons hide cost/latency differences and vendor dependencies, and that vendor-reported numbers need independent verification. Finally, it summarizes how to use Fugu via an OpenAI-compatible API (including sample code, pricing, and open reimplementations), and closes with practical guidance on when to use Fugu (multi-step, messy workflows and robustness needs) versus when a single-model setup is sufficient.
Read the full blog for free on Medium.
Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.
Published via Towards AI
Towards AI Academy
We Build Enterprise-Grade AI. We'll Teach You to Master It Too.
15 engineers. 100,000+ students. Towards AI Academy teaches what actually survives production.
Start free — no commitment:
→ 6-Day Agentic AI Engineering Email Guide — one practical lesson per day
→ Agents Architecture Cheatsheet — 3 years of architecture decisions in 6 pages
Our courses:
→ AI Engineering Certification — 90+ lessons from project selection to deployed product. The most comprehensive practical LLM course out there.
→ Agent Engineering Course — Hands on with production agent architectures, memory, routing, and eval frameworks — built from real enterprise engagements.
→ AI for Work — Understand, evaluate, and apply AI for complex work tasks.
Note: Article content contains the views of the contributing authors and not Towards AI.
Related posts
Recent Posts
