Cost per token is only half the story — the shape of your workload decides which model is actually cheapest. A dated pricing table and a routing framework for matching GPT-4o, Claude, and Gemini to input-heavy, output-heavy, and long-context agent workloads, plus the cost levers that move the number.