AI agents use 600× more energy than chat prompts, study finds

Google and OpenAI have long publicized low energy-per-query figures—0.24 watt-hours for Gemini and 0.34 watt-hours for ChatGPT—to reassure users about AI's environmental impact. However, those benchmarks capture only simple text prompts and ignore the far more complex workloads driving actual AI adoption: reasoning models, multi-agent systems, code generation, and multimodal processing. Hausfather's eight-week personal audit exposes the chasm between those marketing claims and measured reality. By logging every model call and token count from Claude Code, he found that a single "prompt" on his end triggered an average of twelve model calls and processed 2.9 million tokens—compared to roughly a thousand tokens for a typical chat exchange. Crucially, 96 percent of those tokens were cache reads: at each of the 14,000 steps, the agent re-read its entire accumulated context, meaning the output Hausfather actually saw on screen accounted for just 0.4 percent of all processed tokens.

The practical implications are significant. A single day of agent use consumed 3.0 kWh on average—equivalent to two refrigerators running for 24 hours—and his most intensive day peaked at 11 kWh, more than one-third of an average U.S. household's daily electricity draw. Yet Hausfather's framing avoids both alarmism and guilt: he notes that scaling his annual heavy agent use to his full carbon footprint represents only about two percent of an average American's yearly emissions. The real concern is trajectory. AI labs are already planning to scale agent systems that run autonomously for weeks, months, or longer—a shift that Hausfather warns could drive an exponential jump in energy consumption. His conclusion sidesteps calls for individual restraint and instead identifies the carbon intensity of the electricity grid as the decisive lever: a 90 percent emissions reduction is possible if workloads shift to clean power. That reframes the problem from consumer behavior to infrastructure policy, placing responsibility on data centers and energy providers rather than users.