Orchestration design, not model choice, drives agentic AI token costs
arXiv paper isolates the real cost lever in enterprise agentic AI: the orchestration layer (context assembly, tool exposure, turn sequencing, governance) — not the foundation model itself.
Controlled experiment across six models (Claude Sonnet, Gemini 3.1, Gemini Flash, Qwen, GLM, Palmyra) on 22 locked tasks shows token spend per task is dominated by harness design choices, not model capability.
Pattern: falling per-token prices mask rising total spend because longer reasoning traces, more turns, wider tool payloads, and replayed contexts grow faster than task value.