AI agents may use vastly more energy than simple chat prompts

Hausfather’s Claude Code logs covered eight weeks, 1,138 typed prompts, more than 14,000 model calls, and 3.2 billion processed tokens. The Decoder reports that this translated to an estimated 170 kWh of data center electricity, or roughly 150 Wh per input. That is about 600 times the energy of a median chat prompt, showing why “per prompt” comparisons can understate agent-based workloads.

According to the report, each Claude Code prompt triggered an average of twelve model calls and processed an average of 2.9 million tokens. A key reason is that the agent repeatedly re-reads accumulated context, with 96 percent of the processed tokens counted as cache reads. The text visible to the user accounted for only 0.4 percent of all processed tokens, so the apparent output can hide much larger behind-the-scenes computation.

Hausfather’s median Claude Code session consumed about 0.6 kWh, while his average day of use reached 3.0 kWh, with an estimated 11 kWh on his most intensive day. Annualized, the workload would use about 1.1 MWh of data center electricity and produce about 370 kg of CO₂ equivalents based on the average U.S. electricity mix. The Decoder notes that this is slightly more than running an electric clothes dryer for a year and about half of a round-trip economy flight from San Francisco to New York.

Hausfather does not frame the issue mainly as a call for individual guilt or sacrifice. He argues that routing simple tasks to smaller models can help, because they use five to seven times less energy per token than frontier models. But the biggest lever is the carbon intensity of electricity: running the same workload on largely clean power could cut the carbon footprint by about 90 percent.