Hybrid model routing sends each task to the cheapest model that can handle it, cutting inference spend by 40–70% on mixed workloads. This guide covers routing architecture, cost math, and the failure modes that eat the savings — for builders who want control, not hand-holding.