IBM Research published findings from building a model router into agentic systems, reporting results on the AppWorld Test Challenge with a CodeAct agent. Across 417 tasks, Claude Sonnet 4.6 cost $79 total ($0.19/task) while GPT-4.1 cost $155 ($0.37/task), despite GPT-4.1's lower token pricing and Sonnet taking roughly three times as many reasoning steps; the authors attribute the inversion to Sonnet's cheaper cache reads combined with high cache hit rates in agent workloads. They argue routing is an optimization problem rather than a classification one, since task difficulty is often unknowable before execution and routers must simultaneously weigh cost, latency, specialization, reliability and enterprise governance constraints such as data residency and approved model lists.