Input context, not model output, drives 99.2% of agent token costs

An analysis of agent token consumption finds that model output accounts for only about 0.8-0.9% of total usage, while the remaining 99%+ comes from input resent with each call. Measured per task rather than per call, a median task consumed roughly 18 million tokens across about ten chained model calls. The author argues that optimizing the model, prompt, or temperature therefore moves less than 1% of the bill, since cost is determined by how often the context is re-sent. The report describes its figures as reproducible from a frozen data source with a deterministic script. A related finding is that trimming tool schemas from 73 to 22 cut the schema block by 58%, but only about 8.7% of total spend, and that fewer tools did not reliably reduce cost.