The way we measure the cost of AI coding is flawed. When a developer gives an AI agent a task, it doesn't just answer once. It plans, generates code, runs tests, and repeats the process to correct errors. This can result in 10 to 50 times more tokens being used than a basic chat prompt. As AI agents become more useful, they make more calls, and with per-token pricing, the workflows that create the most value can also become the most expensive. This financial pressure is already being felt, with companies like Uber reportedly exhausting its 2026 AI budget by April after expanding its use of Claude Code, and GitHub Copilot moving from flat-rate pricing to metered credits. Gartner predicts that AI coding token costs will surpass average developer salaries by 2028. To address this issue, some companies are implementing usage caps, seat limits, and approvals before developers launch agents. However, these blanket limits can restrict the workflows that deliver the most value and optimize the wrong number. A better approach is to measure the cost per merged pull request, which takes into account the total AI-assisted development cost and the number of changes that pass review and are merged. This metric allows developers to track rework and defects alongside the cost, ensuring that cheaper changes don't conceal lower quality. If the cost per merged pull request falls as AI usage rises without compromising quality, it indicates that developers are converting more consumption into useful software. The key to improving this equation is to reserve expensive inference for tasks that need it and lower the cost of routine work. This can be achieved by matching the model to the task, using a frontier model or an on-premises flagship for difficult planning and novel reasoning, and routing routine work to a more cost-effective model. By making these changes, developers can optimize their AI usage and produce more value from their investment.