AI agents consume roughly 600 times more energy than a standard chat request
Climate scientist Zeke Hausfather has published data showing that autonomous AI systems use far more electricity than the low figures reported by Google and OpenAI suggest.
Over eight weeks, Hausfather logged his use of the programming agent Claude Code. He generated roughly 1,100 inputs, which triggered 14,000 model calls and processed 3.2 billion tokens. This activity consumed about 170 kilowatt-hours of electricity. That averages to around 150 watt-hours per input.
If extended to a full year, this level of intensive use produces carbon dioxide emissions comparable to running an electric clothes dryer. Hausfather identifies the shift to clean energy sources for data centers as the most critical step to cut emissions.
Google and OpenAI have been reassuring users with low-energy consumption figures per AI query. Hausfather’s data tells a different story.
Google says a median Gemini text prompt uses just 0.24 watt-hours, less energy than nine seconds of television. OpenAI CEO Sam Altman estimated the average ChatGPT query at 0.34 Wh, roughly on par with a Google search from 2009.
Both data points were published last year. Even then, it was clear the numbers misrepresent actual AI usage. They do not account for reasoning models that generate many times more tokens from a single chat query, multimodal processing across various formats, multi-agent systems, code generation, scaling across billions of uses, and plenty of other factors. Google’s study particularly downplayed real energy consumption.
Hausfather has now calculated just how big that gap is in an analysis on The Climate Brink. One major blind spot remains: nobody outside AI labs knows the actual energy cost per token of a state-of-the-art model. Keep in mind that Hausfather’s numbers are reasonable estimates, but still just estimates.
They are also based on current usage patterns. AI labs are already working to scale agent-based systems that run autonomously on tasks for days, weeks, or even months. That could drive an exponential jump in energy consumption.
3.2 billion tokens in eight weeks
Hausfather tracked his own usage in detail over eight weeks. Claude Code stores complete local logs of every session, including exact token counts reported by the API for each model call.
His 1,138 typed prompts triggered over 14,000 model calls, an average of twelve per prompt. Each prompt processed an average of 2.9 million tokens. For comparison, a typical chat exchange without reasoning or web searches runs about a thousand tokens.
In total, his Claude Code processed 3.2 billion tokens. Of those, 96 percent were cache reads because at each of the 14,000 steps, the agent re-reads its entire accumulated context. The text Hausfather actually sees on screen, the model’s output, accounts for just 0.4 percent of all processed tokens.
His best estimate for total consumption comes to about 170 kWh of data center electricity over eight weeks, with an uncertainty range of 70 to 330 kWh. Hausfather directly measured only the token counts from his Claude Code logs, and the conversion to electricity values relies on three independent methods with different assumptions.
Per prompt, that works out to roughly 150 Wh, about 600 times as much as a median chat prompt. “A ‘prompt’ is ultimately not a unit of AI use any more than ‘trips’ is a measurement of driving; it’s how far you go that matters,” Hausfather writes.
A single day of agent use draws more power than two refrigerators
Hausfather’s median Claude Code session consumed about 0.6 kWh, fifty times the electricity needed to charge a phone. His average day using Claude Code hit 3.0 kWh, more than the daily draw of two refrigerators.
On his most intensive day, when several parallel agents were grinding through an extensive geodata analysis, he estimates 11 kWh were consumed. That is more than a third of the daily electricity consumption of an average U.S. household, according to Hausfather.
A year of heavy agent use matches a clothes dryer
Scaled to a full year, Hausfather’s agent-based Claude Code usage would consume about 1.1 MWh of data center electricity, roughly one-tenth of what an average U.S. household uses annually. Based on the average U.S. electricity mix, that translates to about 370 kg of CO₂ equivalents per year.
That 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, though Hausfather notes the flight figure covers only direct CO₂ emissions. It amounts to about eight percent of the annual emissions from a typical U.S. gas-powered car and roughly two percent of the average American’s yearly carbon footprint. “This is simultaneously a large emissions source and a relatively modest part of my total carbon footprint,” Hausfather writes.
Clean electricity matters more than cutting back
Hausfather does not argue for sacrifice or guilt. Personal restraint by the small group of heavy users is not going to bend any curves, he says, adding that routing simple tasks to smaller models does make sense, since they use five to seven times less energy per token than frontier models.
But the biggest lever is the carbon intensity of the electricity itself, he argues. If the same workload ran on largely clean power, the carbon footprint would drop by about 90 percent. The problem is that nearly three-quarters of planned on-site power generation for U.S. data centers runs on natural gas.
That is where Hausfather sees an opening. AI companies bring enormous capital and unusual urgency. If that money flows into clean energy, grid expansion, and advanced technologies like geothermal or nuclear power, the AI boom could leave the grid cleaner than it found it.
What it means
For people building software or running analysis, the implication is that the energy cost is hidden in the background. A single session can consume as much power as household appliances. The focus needs to shift from counting individual prompts to managing the total volume of tokens and the source of the electricity powering the servers.
