Research Summary: How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks

Research into AI agent token consumption in coding tasks reveals that these agents are significantly more expensive than traditional code reasoning, consuming tokens at a rate 1000 times higher. This study, the first systematic analysis of its kind, examines token usage patterns across eight leading LLMs on the SWE-bench Verified dataset and assesses their predictive capabilities for token costs. Understanding and managing this consumption is critical given the rapid adoption of AI agents in complex workflows.

Why it matters

The escalating token consumption by AI agents in complex workflows presents a significant operational cost factor that can impact budget allocation and the economic viability of AI deployments. Understanding where tokens are spent, identifying more efficient models, and improving cost prediction are critical for sustainable scaling and strategic investment in AI capabilities.

Key insights

• AI agents integrated into complex human workflows are driving a rapid increase in Large Language Model (LLM) token consumption.
• Agentic tasks are uniquely expensive, consuming approximately 1000 times more tokens compared to code reasoning tasks.
• The study systematically analyzed token consumption patterns in agentic coding tasks across eight frontier LLMs on the SWE-bench Verified benchmark.
• Models' ability to predict their own token costs prior to task execution was evaluated.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2604.22750