AI Agents Consume About 600 Times More Energy Than Simple Chat Prompts

According to Google, a median Gemini text prompt uses only 0.24 watt-hours, while OpenAI CEO Sam Altman estimated a ChatGPT request at 0.34 Wh. Climate scientist Zeke Hausfather argues these figures misrepresent real AI usage, as they ignore reasoning models, multimodal processing, and agentic systems. In his analysis on The Climate Brink, he logged eight weeks of his own Claude Code usage, which stored detailed local logs with exact token counts. His 1,138 typed prompts triggered over 14,000 model calls, averaging twelve per prompt, each processing 2.9 million tokens on average, totaling 3.2 billion tokens, with 96% being cache reads. His best estimate for total energy consumption was around 170 kWh (with uncertainty 70-330 kWh), translating to about 150 Wh per prompt, approximately 600 times that of a median chat prompt. A median session consumed 0.6 kWh, his average day 3.0 kWh, and his most intensive day 11 kWh. Scaled annually, this would be 1.1 MWh, causing around 370 kg CO2-equivalent per year, more than an electric clothes dryer (262 kg) and about half a round-trip flight from San Francisco to New York. Hausfather does not argue for abstinence; he suggests delegating simple tasks to smaller models and, more importantly, using clean energy, as it could reduce the footprint by 90%. He notes that nearly three-quarters of planned U.S. data center power generation relies on natural gas, but sees an opportunity for AI companies to invest in clean energy and grid improvements.