Sakana AI's 7B model manages GPT-5 and Claude

What if frontier language models like GPT-5 and Claude did not have to work in isolation or follow rigid, human-made scripts?

Sakana AI took inspiration from how humans build companies: when work gets hard, we do not expect one genius to do everything. We build a team, assign roles, and delegate.

In a paper accepted at ICLR 2026, the Tokyo-based lab unveiled Conductor, a compact 7B model. Conductor does not write code or solve problems directly. Instead, it acts as an executive manager orchestrating frontier models like GPT-5, Gemini, Claude, and open-source models.

A 7B manager that writes dynamic prompts

Traditional multi-agent setups usually fall into two traps: rigid, handcrafted pipelines that break easily, or naive routers that merely pick one model to answer everything.

Sakana trained Conductor using pure reinforcement learning to output collaborative workflows in natural language.

For any incoming problem, the 7B orchestrator decides three things: which agent to call, what specific subtask to hand off as an expert meta-prompt engineer, and what prior messages belong in their context window.

Through reward maximization, the manager learned to adapt to task difficulty on the fly. It 1-shots easy factual questions to save compute, but autonomously spins up a full planner-executor-verifier pipeline when facing brutal programming challenges.

It also learned recursive test-time scaling. By selecting itself as a worker, Conductor reads what its team produced, spots flaws, and spins up a corrective workflow right away.

Beating solo frontier models on a budget

The benchmark results justify the management layer.

The 7B Conductor outperformed every single frontier model in its pool when running solo, setting records on LiveCodeBench at 83.9% and GPQA-Diamond at 87.5% at publication time.

Crucially, it also outpaced resource-heavy baselines like Mixture-of-Agents at a fraction of the cost, precisely because it does not broadcast every question to every worker.

Together with Sakana's recently announced TRINITY framework, this foundational research powers their new multi-agent product, Sakana Fugu.

Why it matters

We think Conductor points to where enterprise AI is actually heading: model-agnostic management instead of betting on a single monolithic system.

Most teams building agent pipelines today spend weeks hand-crafting prompts and conditional branches. Sakana shows that workflow orchestration can be learned entirely by a small, cheap 7B model using natural language as the control bus.

For developers, this turns heterogeneous model routing from a maintenance headache into an automated, cost-effective layer. You do not have to choose exclusively between Claude and OpenAI.

The trade-off, however, is inference latency and cumulative API fragility. When multiple frontier models are invoked in a recursive loop, response times inevitably stretch, and an error by the meta-prompt manager can still derail the entire pipeline.

Still, shifting compute scaling from brute-force pretraining to dynamic, inference-time team management looks like one of the smartest engineering bets Sakana has made.