77.27% average across major benchmarks. That's the score Sakana AI's RL Conductor just posted — a 7-billion-parameter model that doesn't answer questions itself. Instead, it orchestrates GPT-5, Claude, and Gemini to produce better results than any of them can alone. A model 20x smaller than the systems it's conducting just beat them at their own game.
The Problem
The AI industry has been running a brute-force race. Bigger models. More parameters. More training data. More compute. GPT-5, Claude Opus, Gemini Ultra — each generation demands exponentially more resources to squeeze out marginal improvements.
This is expensive. Frontier model training runs now cost hundreds of millions of dollars. Inference for these models isn't cheap either. And the returns are diminishing — each new generation delivers smaller jumps in capability while consuming dramatically more resources.
But here's the thing most teams miss: the frontier models are already good enough individually. The bottleneck isn't model quality. It's knowing which model to use for which part of a task, and how to combine their outputs.
Most organizations pick one model and use it for everything. That's like hiring one consultant and asking them to be your lawyer, accountant, and marketing director. You get mediocre everything instead of excellent anything.
The Solution
Sakana AI trained a 7B-parameter conductor model using reinforcement learning to do one thing: decide which frontier model should handle each sub-task, then synthesize the results.
The approach is deceptively simple:
- The conductor doesn't generate answers. It generates routing decisions. Given a problem, it breaks it into sub-tasks and assigns each to the best-fit frontier model.
- Trained via RL, not supervised learning. The conductor learned through trial and error which routing strategies produce the best outcomes. No hand-labeled examples needed.
- Works with any frontier model. The conductor is model-agnostic. Swap in new models as they're released without retraining the whole system.
Think of it like an expert project manager who doesn't write code but knows exactly which senior engineer should tackle each part of the architecture. The manager is "small" compared to the engineers. But the output is better than any single engineer working alone.
This is the orchestration paradigm: small, specialized models coordinating large, general-purpose ones. It's cheaper, faster, and — as the benchmarks show — more effective.
Benchmarks
The RL Conductor's numbers are striking:
- 77.27% average across all benchmarks — higher than any individual frontier model
- 93.3% on AIME25 (mathematical reasoning)
- 87.5% on GPQA-Diamond (graduate-level science QA)
- 83.93% on LiveCodeBench (real-world coding tasks)
Caveats worth noting:
- The conductor adds latency overhead — routing decisions and multi-model calls take more wall-clock time than a single model inference
- API costs multiply — you're calling multiple frontier models per task instead of one (though the per-call cost is lower since each model handles a smaller sub-task)
- The benchmarks are academic — real-world production tasks may show different patterns
- Sakana hasn't open-sourced the training code yet, so independent reproduction is pending
Even with those caveats, the signal is clear. Coordinated small models beat solo large models.
Impact
This shifts the economics of AI deployment fundamentally.
For enterprises: You don't need to bet on one model provider. An orchestration layer lets you use the best model for each task, avoiding vendor lock-in while improving results. A conductor model is cheap to run — 7B parameters means you can host it locally for pennies.
For AI teams: Training a conductor is vastly cheaper than training a frontier model. Sakana's approach suggests a path where smaller companies can compete not by building bigger models, but by building smarter orchestration. This is the "last mile" of AI — not raw capability, but intelligent routing and combination.
For the industry: Allen AI's EMO model reinforces this trend from a different angle. Their MoE (Mixture of Experts) architecture lets you use just 12.5% of experts for a specific task while retaining near-full performance. Standard MoE models degrade severely when you drop experts. EMO doesn't. Modularity and selective deployment are the new frontier.
The combined signal: AI efficiency is moving from "build bigger" to "deploy smarter." The companies that figure out orchestration first will have a structural cost advantage over those still paying for single-model brute force.
The Bottom Line
The future of AI isn't one model to rule them all. It's a small, sharp conductor making the big models dance.
If you're building AI products and you're not thinking about orchestration, you're leaving performance on the table and money in someone else's pocket. The 7B model that beats GPT-5 isn't a curiosity — it's a preview of how the next generation of AI systems will actually work.