A joint study by Stanford, MIT, and other institutions has unveiled the high token consumption of AI agents in coding tasks, revealing that these agents can burn millions of tokens while attempting to fix code bugs. The research, published in April 2026, highlights that the cost of having an AI agent write code is approximately 1,000 times higher than standard AI conversations due to the extensive 'reading' of code required. This involves feeding the model with project context, operation logs, and error messages, leading to exponential growth in input tokens. The study also found significant variability in costs, with the same task potentially costing twice as much on different runs. Additionally, the research identified that some models, like GPT-5, are more token-efficient than others, impacting financial outcomes in enterprise applications. The findings suggest that current AI models lack 'stop-loss awareness,' often consuming more tokens on unsolvable tasks. The study calls for the development of budget-aware policies to manage token consumption effectively, as unpredictable costs challenge the sustainability of subscription pricing models in AI agent scenarios.