University of Michigan study finds AI agents have unpredictable, soaring token costs
Updated
Updated · ZDNet · May 5
University of Michigan study finds AI agents have unpredictable, soaring token costs
12 articles · Updated · ZDNet · May 5
Researchers said agentic coding tasks can use about 3,500 times more tokens than simple ChatGPT-style prompts, with the same model sometimes costing twice as much on identical runs.
The study, involving Stanford, MIT, Microsoft, DeepMind and others, found models systematically underestimate token needs and that higher spending often fails to improve accuracy.
Input tokens, especially repeated cache reads and retrieved context, drove most costs, underscoring calls for vendors including OpenAI, Google and Anthropic to offer clearer pricing and performance guarantees.
With token usage so unpredictable, will outcome-based pricing models become the industry standard, or are there deeper challenges to be solved?
Could hidden and unpredictable AI agent costs derail the promised productivity gains and ROI for enterprises betting on automation?
The 1000x Token Consumption Crisis: Unmasking the Hidden Costs of AI Coding Agents
Overview
A landmark April 2026 study revealed that AI coding agents consume about 1,000 times more tokens than typical conversational AI, driven by repeated ingestion of massive project data and unproductive exploration loops. This token usage is highly inconsistent, with up to 30-fold variation for identical tasks, and throwing more tokens rarely improves results. These inefficiencies have made subscription pricing models untenable, causing financial turmoil for enterprises and forcing a shift to pay-as-you-go billing. CFOs now face cost volatility and budgeting challenges. In response, the industry is developing practical mitigation strategies and pushing for greater transparency, smarter context management, and human oversight to ensure AI tools become more efficient, predictable, and economically sustainable.