Orchestrator-8B

ToolOrchestra trains an 8B controller that alternates between reasoning and tool calls over multiple turns, choosing among web search, code interpreters, specialized math or coding models, and generalist LLMs such as GPT-5, Llama-Nemotron-Ultra, and Claude Opus 4.1. It uses reinforcement learning with outcome, efficiency, and preference rewards so the orchestrator learns not only to solve tasks, but to do so cheaply and in line with user tool preferences. On benchmarks like Humanity’s Last Exam, τ²-Bench, and FRAMES, Orchestrator-8B reaches higher accuracy than GPT-5 while using around a third of the cost and showing robust generalization to unseen tools, suggesting that a lightweight tool router can be more cost-effective than a single giant model.