Where did all my Claude Code tokens go?

The text discusses the challenges and insights surrounding the evaluation of AI coding agents like Claude Code, particularly how their costs and productivity are assessed. It highlights the industry's reliance on subjective feelings rather than clear metrics to evaluate the impact of AI-generated code on productivity, citing a report where many engineers identify this as a significant challenge. The author conducts an experiment using three different approaches to build an app, comparing the costs and productivity of each method. The experiment reveals that while test-driven development appears cost-effective initially, the volume of work each method handles complicates direct cost comparisons. The analysis delves into the importance of cache hit rates and the cost implications of AI agents' decision-making processes during sessions. The text underscores the significance of understanding session dynamics and context management to control AI usage costs effectively, proposing that real-time data analysis rather than mere intuition can optimize AI's financial and operational benefits.