A Climate Scientist Logged Eight Weeks of AI Agent Use. It Drew 600 Times More Power Per Prompt Than a Chat Message
Zeke Hausfather tracked 3.2 billion tokens through Claude Code and estimated the electricity behind them. The per-prompt figures the AI labs publish turn out to describe a kind of use almost nobody does anymore.
For a year the AI industry has answered questions about its power use with a reassuringly small number. Google says the median Gemini text prompt uses 0.24 watt hours, less than nine seconds of television. Sam Altman put the average ChatGPT query at 0.34 watt hours. Climate scientist Zeke Hausfather has now measured what happens when you use AI the way a growing number of people actually do, and the number is roughly 600 times bigger.
Hausfather spent eight weeks logging his use of Claude Code, a coding agent that works through tasks by itself rather than answering one question at a time. The tool keeps local logs with the exact token count the API reports for every call, so the counting is solid. He typed 1,138 prompts. Those triggered more than 14,000 model calls, about twelve per prompt, and processed 3.2 billion tokens. A token is roughly a word fragment, and an ordinary chat exchange runs about a thousand of them. His average prompt ran 2.9 million.
The reason for that gap is worth understanding. Ninety six percent of those tokens were cache reads: at every one of the 14,000 steps, the agent re-reads its whole accumulated context to remember what it is doing. The output he actually saw on screen was 0.4 percent of everything processed. An agent spends almost all its effort re-reading its own notes.
Converting tokens to electricity is the soft part, and Hausfather says so plainly. Nobody outside the labs knows the true energy cost per token of a frontier model, so he used three independent methods. His best estimate is about 170 kilowatt hours over eight weeks, with a range of 70 to 330. That works out to roughly 150 watt hours per prompt. A typical day of agent use came to 3.0 kWh, more than two refrigerators draw.
Scaled to a year, that habit would produce around 370 kg of CO2 on the average US grid. Which is where it gets less alarming than the headline suggests: that is slightly more than running an electric clothes dryer for a year, and about half a round trip economy flight from San Francisco to New York. Hausfather’s own summary is that it is simultaneously a large emissions source and a modest part of his total footprint.
What this means for you: if you use chatbots the ordinary way, a few prompts a day, your AI footprint is genuinely tiny, around 0.3 kg of CO2 a year, and guilt is misplaced. If you have started running agents that work unattended for an hour, you have quietly moved several orders of magnitude up the scale, and that is worth knowing. The practical lever is not abstinence but model choice, since smaller models use five to seven times less energy per token, so sending simple tasks to a smaller one is real savings. Hausfather’s conclusion points past individuals entirely: nearly three quarters of planned on-site power generation for US data centres runs on natural gas, and the same workload on clean electricity would cut the footprint by about 90 percent. A fair caveat on all of it: these are careful estimates, not meter readings.
Sources
Source: https://www.theclimatebrink.com/p/the-real-energy-use-of-agentic-ai
Backflip AI Turns 3D Scans Into Editable CAD Models, a Job That Usually Eats Hours
Most factories have digital models for less than one percent of their parts. A new model converts a scan into real, editable CAD geometry rather than a dead mesh, and runs as an Autodesk Fusion add-in.
