Routing, Cascades, and User Choice for LLMs

A new ICLR 2026 paper uses game theory to analyze how LLM providers route user tasks between cheaper 'standard' and more expensive 'reasoning' models. The research reveals a significant misalignment between provider and user goals. In most cases, optimal routing is a simple static policy. Alarmingly, the study shows providers are sometimes incentivized to intentionally throttle model latency to minimize their own costs, which directly depresses user utility and satisfaction.