Token costs are the smallest part of running an AI agent, study finds

A technical analysis published on the Dev Community blog argues that API token fees are often the least significant expense when running large language model agents in production. The post identifies six cost axes that operators should track: token usage, latency, orchestration infrastructure, third-party tool-call fees, human-in-the-loop review time, and idle polling overhead. Because cloud providers only invoice for token consumption, the other five cost categories tend to go unbudgeted and untracked, spread invisibly across cloud bills, staff calendars, and latency dashboards. The author notes that human review time is typically one to two orders of magnitude more expensive per minute than compute, and that idle worker processes can cost more than active inference when throughput is low. A sample cost model accompanying the post lets teams substitute their own figures to identify which axis actually dominates their specific workload.