Electricity use of AI coding agents
Summary
The post estimates the energy cost of using Claude Code for AI coding sessions by extending median-query energy models to session-level workloads. It introduces napkin-math approximations for input, output, and cache-related tokens, and presents rough Wh per million tokens figures, concluding that Claude Code sessions are orders of magnitude more energy-intensive than typical queries. The author also provides session logs, per-token energy reasoning, and a cost snapshot to contextualize environmental impact and personal spend.