Forget the polite fiction that a single AI query sips less power than a few seconds of TV. When climate scientist Zeke Hausfather started logging his actual use of Anthropic’s Claude Code, the numbers told a radically different story. Over eight weeks, his 1,138 typed prompts triggered more than 14,000 model calls and processed 3.2 billion tokens. The damage: roughly 170 kilowatt-hours of electricity, or about 150 watt-hours per input.
That’s not a rounding error. It’s roughly 600 times the energy Google cites for a basic Gemini text prompt. The core issue isn’t just more use—it’s the architecture of agents themselves. At each of those 14,000 steps, the model re-reads its entire accumulated context. Fully 96 percent of the tokens Hausfather’s agent processed were cache reads, invisible churn that never produced a single word of output for him to see. His median day of coding burned 3.0 kWh, more electricity than running two refrigerators.
There’s a crucial caveat in this data that should keep both AI boosters and critics humble: nobody outside the labs actually knows the real energy cost per token of a frontier model. Hausfather’s figures are a best-guess triangulation using three independent estimation methods, with an uncertainty range from 70 to 330 kWh over that eight-week period. The ground truth is locked inside private data centers.
Extrapolated to a full year of heavy agent use, the emissions hit about 370 kg of CO₂—slightly more than your clothes dryer, and about half a round-trip flight from San Francisco to New York. Hausfather calls it simultaneously a large emissions source and a relatively modest slice of his personal carbon footprint. The bigger worry is what happens if AI labs succeed in their goal of deploying autonomous agents that grind on tasks for weeks or months at a stretch. That’s the scenario where efficiency gains get swamped by sheer scale, which is why Hausfather keeps coming back to one lever that actually matters: cleaning up the grid itself.
