Sakana trains 7B model to orchestrate GPT, Claude, Gemini
Sakana AI has developed RL Conductor, a 7-billion-parameter language model trained via reinforcement learning to automatically orchestrate calls to larger frontier models like GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro. Rather than relying on hard-coded routing logic, the model learns to dynamically analyze inputs, distribute work among specialized agents, and coordinate responses. The approach achieves state-of-the-art results on reasoning and coding benchmarks while reducing API costs and call volume compared to both individual frontier models and manually designed multi-agent systems.
TL;DR
- Sakana AI's RL Conductor is a 7B model trained to automatically route tasks to a pool of larger LLMs based on input characteristics and task requirements
- The system outperforms individual frontier models and hand-designed multi-agent pipelines on reasoning and coding benchmarks while cutting costs and API calls
- RL Conductor learns orchestration strategies through reinforcement learning rather than human design, enabling it to adapt to shifting query distributions and heterogeneous user demands
- The technology powers Fugu, Sakana AI's commercial multi-agent orchestration service, addressing a core limitation of rigid frameworks like LangChain
Why It Matters
This work directly challenges the assumption that larger models always perform better. By training a smaller model to intelligently delegate to specialized larger models, Sakana demonstrates that orchestration itself is a learnable skill. This has implications for how teams build production AI systems, suggesting that the future of agentic AI may depend less on scaling individual models and more on intelligent coordination across diverse model pools.
Business Impact
For operators and founders building multi-model systems, this approach offers a path to better performance at lower cost. Hard-coded routing breaks in production when user queries shift or diversify. An adaptive orchestrator that learns which model to use for which task could reduce both infrastructure spend and latency, making it economically viable to maintain a diverse pool of specialized models rather than defaulting to a single large model.
Key Implications
- Smaller models can add significant value by learning to coordinate larger ones, potentially shifting investment away from pure scale and toward orchestration logic
- Reinforcement learning can discover orchestration strategies that humans would struggle to hand-code, including iterative refinement and dynamic communication topologies tailored per query
- The brittleness of hard-coded agentic frameworks is a real production bottleneck, and automated adaptation to heterogeneous workloads is becoming a competitive necessity
What to Watch
Monitor whether RL Conductor's approach generalizes across different model pools and domains, and whether other labs adopt similar reinforcement learning-based orchestration. Also watch for adoption metrics on Fugu and whether this model of intelligent routing becomes a standard layer in production AI stacks, potentially shifting how teams architect multi-model systems.
Subscribe to the newsletter
The latest stories and analysis, delivered to your inbox.
Free. No spam. Unsubscribe any time.
