I paired seven months of coding-agent transcripts with a token inventory (morkeeth.substack.com)

In a revealing case study, a developer analyzed the token usage from seven months of interactions with AI coding agents Claude Code and Codex, totaling an impressive $33,807 at list price. The investigation highlighted that 42% of tokens were consumed by agents managing other agents, while only 15% were applied to user-facing applications. Notably, a single hour resulted in 284 million tokens being logged, with an astonishing 96.66% of those dedicated to cache reads—essentially the agents re-reading information rather than generating new output. This raises significant questions about efficiency and the practical implications of token economics in AI workflows. The findings underscore critical considerations for the AI/ML community regarding resource allocation and operational efficiency in machine learning models. The author provided concrete measures to optimize token usage, including minimizing startup context and running sessions with a single job focus, which could help developers reduce unnecessary costs and improve productivity. Furthermore, by illustrating how most tokens were spent managing internal agent communications rather than delivering external results, the study emphasizes the need for a more strategic approach in designing and implementing AI systems.