How Do Coding Agents Spend Your Money? Analyzing and Predicting Token Consumptions in Agentic Coding Tasks

ICLR 2026 Conference Submission20681 Authors

Keywords: AI Agent, Coding Agent

Abstract: AI agents offer substantial opportunities to boost human productivity across many settings. However, their use in complex workflows also drives rapid growth in LLM token consumption. When agents are deployed on tasks that can require millions of tokens, a natural question arises: where does token consumption come from in agentic coding tasks, and can we predict how many tokens a task will require? In this paper, we present the first empirical analysis of agent token consumption patterns using agent trajectories on SWE-bench, and we further explore the possibility of predicting token costs at the beginning of task execution. We find that (1) more complex tasks tend to consume more tokens, yet token usage also exhibits large variance across runs (some runs use up to 10$\times$ more tokens than others); (2) unlike chat and reasoning tasks, input tokens dominate overall consumption and cost, even with token caching; and (3) while predicting total token consumption before execution is very challenging (Pearson’s $r<0.15$), predicting output-token amounts and the range of total consumption appears practical and reasonably accurate. Understanding and predicting agentic token consumption is a key step toward transparent and reliable agent pricing. Our study provides important empirical evidence on the inherent challenges of token consumption prediction and could inspire new studies in this direction.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 20681