My take: Sakana trains 7B model to orchestrate GPT, Claude, Gemini

The Conductor Is Smarter Than the Orchestra

A small model telling GPT-5 what to do. That's the idea, and it's a good one.

Sakana's RL Conductor flips the usual assumption that bigger always means better. A lean 7B model, trained through reinforcement learning, learns to route tasks to the right frontier model at the right time. Less waste. Lower costs. Better results on benchmarks than using any single frontier model alone.

This matters because the AI cost problem is real and getting worse. Every unnecessary API call to GPT-5 or Gemini 2.5 Pro is money burned. Most teams are still duct-taping routing logic together by hand, and that doesn't scale. A learned orchestrator that figures this out automatically is the kind of infrastructure the industry actually needs right now.

The deeper signal here is about where intelligence should live. Not every problem needs a frontier model. Sometimes you need a traffic controller more than you need a genius, and a smart small model in that role can outperform a powerful model working alone and confused.

Watch this space. Orchestration is quietly becoming the most valuable layer in the stack.