I ran Claude Code most of last week without once wondering what it costs in watts. Hausfather’s number — 150 Wh per prompt against Altman’s 0.34 Wh for a chatbot request — reframes the bill. His 1,138 prompts became 14,000 model requests and 3.2 billion tokens. That fan-out is the story: I type one line, the agent decides how much machine to spend.
So I am doing what he actually recommends, which is model selection, not abstinence. Simple tasks on small models run five to seven times less energy per token than always reaching for the frontier model. My boat-data scripts and Astro rebuilds do not need the big one.
The story — Climate researcher Zeke Hausfather, of Stripe and Berkeley Earth, measured his own Claude Code use over eight weeks: 1,138 prompts, 14,000+ model requests, 3.2 billion tokens, roughly 170 kWh (range 70–330). That averages 150 Wh per prompt, versus the 0.34 Wh Sam Altman once cited for a ChatGPT request. Annualized: 1.1 MWh and about 370 kg CO₂ — more than a clothes dryer, less than a California EV. (Source)