AI agents consume far more energy than simple chat prompts
Climate scientist Zeke Hausfather tracked eight weeks of his work with Claude Code and estimated the electricity used by the agent. His 1,138 typed prompts triggered more than 14,000 model calls and processed 3.2 billion tokens. The estimated data-center consumption was about 170 kWh, with a range of 70 to 330 kWh. That equals roughly 150 Wh per prompt, or about 600 times the 0.24 Wh reported for a median Gemini text prompt.
Hausfather’s figures are estimates because the conversion from token counts to electricity depends on assumptions, and the energy cost per token of leading models is not publicly known. Most of the processed tokens—96 percent—were cache reads as the agent repeatedly reloaded its accumulated context. A typical Claude Code day used 3.0 kWh on average, while the most intensive day reached 11 kWh.
If maintained for a year, the usage would require about 1.1 MWh and generate an estimated 370 kg of CO₂ equivalents. Hausfather says cleaner electricity for data centers would have the greatest effect: running the same workload on mostly clean power could reduce emissions by about 90 percent. He also notes that smaller models use five to seven times less energy per token than frontier models.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.