Researchers have developed a common measurement protocol for evaluating four open-source model routers across various tasks and benchmarks. The study compares the performance of these routers on four different benchmarks, using a locked matrix of candidate outcomes. The results show that three routers consistently assign high or low tiers to tasks, while one router varies its assignments based on prompt content. However, this variation does not necessarily lead to better performance, as the best-performing router is actually the one with the most consistent tier assignments. --- Why it matters: This study matters because it provides a framework for comparing and evaluating model routers, which are increasingly used in agentic systems. The findings have implications for the development of more effective routing protocols, particularly in scenarios where task-specific targeting is not necessary. Source: https://arxiv.org/abs/2608.14641