The primary driver of high token costs in AI agents is input token consumption, not output, because each reasoning cycle requires reprocessing the full trajectory of past actions.

It's intuitive to assume that the cost comes from the agent's long, detailed responses. However, detailed analysis reveals the opposite: the vast majority of tokens consumed are input tokens. This happens because, at every step of the agent's loop—before it even decides what action to take—it must reread the complete history of its interaction. This includes every prior plan, every tool call made, every observation recorded, and every self-reflection generated. As this history grows longer with each iteration, the context window becomes saturated with repetitive data. The agent isn't paying for what it says; it's paying to constantly relearn what it has already done, making the input side of each model call the dominant cost factor in autonomous workflows.