Why your AI bill tripled while token prices fell 75%

This post identifies that agentic AI costs are primarily a control problem due to extended reasoning loops, multi-turn dialogues, and multi-hop data retrievals, leading to increased token consumption. It highlights the lack of per-task cost telemetry and observability as a key issue, contrasting with the falling token prices. The post proposes capturing five fields per agent run (tokens in, tokens out, model used, retry count, task completion) to compute a real cost per completed task and suggests routing traffic through a single gateway for unified cost data. It advocates for measuring the task, not the token, by focusing on metrics like cost per completed task, context efficiency, quality-adjusted efficiency, and business value multiple.