One Coordinator, Many AI Brains Working Together
OpenFugu lets a tiny routing model decide which frontier AI — GPT, Claude, or Gemini — handles each part of your task, and it's free to run yourself.
The idea behind it
Imagine you're running a project and you have three specialists on call — one brilliant at deep thinking, one fast at execution, one good at catching mistakes. Instead of picking just one and hoping for the best, you had a smart coordinator who knew exactly who to hand each piece of work to.
That's essentially what a company called Sakana AI built last month. They called it Fugu — a system that sits above GPT, Claude, and Gemini and routes each part of your question to whichever model is best suited for that moment. A tiny coordinator brain (much cheaper to run than the big models) learned, through a kind of trial-and-error training, how to assign roles: Thinker, Worker, Verifier.
Within days, the open-source community reverse-engineered the whole thing from the academic papers and released OpenFugu — a free, self-hostable version of the same idea. Early tests suggest it outperforms any single model used alone by a significant margin.
It's also landed in a small controversy — some researchers publicly questioned Sakana's benchmark numbers — which makes OpenFugu an interesting way to test the concept yourself, without paying for the closed original.
Why this matters for your business
Right now, most businesses pick one AI tool and use it for everything. This is a glimpse at how things might work soon: a quiet coordinator in the background, matching the right model to the right job, automatically.
You don't need to do anything with this today. But it's worth knowing this is where things are heading.
Words worth knowing
Frontier model — The biggest, most capable AI models available right now: GPT-4o, Claude, Gemini. Think of them as the top specialists.
Orchestrator — A system that manages other systems. Like a project manager who never sleeps and doesn't charge by the hour.
Open-source — Software whose inner workings are publicly available. Anyone can inspect it, use it, or build on it — no licence fee.
Benchmark — A standardised test used to compare how well different AI models perform on specific tasks. Like a standardised exam for AIs.