ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

Large language models are powerful generalists, yet solving deep and complex problems such as those of the Humanity’s Last Exam (HLE) remains both conceptually challenging and computationally expensive. We show that small orchestrators managing other models and a variety of tools are able to both push the upper bound of intelligence and improve efficiency in solving difficult agentic tasks. We introduce ToolOrchestra, a method for training small orchestrators that coordinate the use of intelligent tools. ToolOrchestra makes explicit use of reinforcement learning with outcome-, efficiency-, and user-preference-aware rewards. Using ToolOrchestra, we produce Orchestrator, an 8B model that achieves higher accuracy at lower cost than previous tool-use agents while aligning with user preferences on which tools are to be used for a given query. On HLE, Orchestrator achieves a score of 37.1%, outperforming GPT-5 (35.1%) while being 2.5x more efficient. On $\tau ^2$-Bench and FRAMES, Orchestrator surpasses GPT-5 by a wide margin while using only about 30% of the cost. Extensive analysis shows that Orchestrator achieves the best trade-off between performance and cost under multiple metrics, and generalizes robustly to previously unseen tools. These results demonstrate that composing diverse tools with a lightweight orchestration model is both more efficient and more effective than existing methods, paving the way for practical and scalable tool-augmented reasoning systems. These results demonstrate that orchestrating diverse tools with lightweight agents is not only more efficient, but also more effective, paving the way for practical and scalable tool-augmented reasoning systems.

This paper introduces ToolOrchestra, a method for training a small "manager" AI that, instead of solving hard problems on its own, decides which other AI models and tools to call on and coordinates them to reach an answer. Trained with reinforcement learning that rewards correct answers, low cost, and respect for user preferences about which tools to use, the resulting 8-billion-parameter "Orchestrator" punches well above its weight: on the notoriously difficult Humanity's Last Exam benchmark it scored 37.1%, beating the far larger GPT-5 (35.1%) while being about 2.5 times more cost-efficient, and on other benchmarks it outperformed GPT-5 by a wide margin at roughly a third of the cost while also adapting well to tools it had never seen—suggesting that a lightweight coordinator paired with diverse specialized tools can be both cheaper and more capable than relying on a single enormous model.