How to reduce token usage in Claude Code, Codex, and Cursor (www.avanderlee.com)

Recent analysis of token usage in AI tools like Claude Code, Codex, and Cursor revealed that a staggering 94.2% of tokens are consumed by cache reads, as agents repeatedly re-read context rather than producing new outputs. The author observed that significant inefficiencies stem from full-file reads, duplicate skill loads, and unchecked build outputs. By implementing global modifications through a hook script, they were able to enhance token efficiency across multiple projects without altering individual AGENTS.md files. This approach emphasizes the importance of controlling context management for optimizing performance. For the AI/ML community, this research serves as a crucial reminder of how nuanced agent memory and context handling can impact operational costs. By refining how agents manage and access context—such as summarizing build logs or filtering commands—developers can drastically reduce token usage and improve processing speed. The findings underscore the need for tailored strategies to manage agent environments effectively, ultimately leading to more sustainable and cost-effective AI application development.