Claude Code: The Colossal Energy of AI Agents Unveiled

Claude Code: The Colossal Energy Consumption of AI Agents Revealed

AI agents consume about 600 times more energy than a simple chat request.

Climatologist Zeke Hausfather demonstrates in a new analysis that AI agents consume significantly more energy than the energy consumption figures per request reported by Google and OpenAI suggest.

Over an eight-week period, while tracking his use of the programming agent Claude Code, approximately 1,100 inputs generated over 14,000 model calls and 3.2 billion tokens, consuming around 170 kilowatt-hours of electricity, or about 150 watt-hours per input.

Extended over a full year, this level of intensive usage produces CO₂ emissions comparable to those of an electric clothes dryer. Hausfather considers the shift to clean energy sources for data centers as the most critical lever for reducing emissions.

Google and OpenAI have reassured users with figures of low energy consumption per AI request. Zeke Hausfather's data from Claude Code tells a different story.

Google states that a median text request from Gemini uses only 0.24 watt-hours (Wh), less energy than nine seconds of television. OpenAI CEO Sam Altman estimated the average request for ChatGPT at 0.34 Wh, roughly equivalent to a Google search from 2009.

These two figures were published last year. Even at that time, it was clear that these numbers did not represent the actual usage of AI. They do not account for reasoning patterns that generate numerous tokens from a single chat request, multimodal processing across various formats, multi-agent systems, code generation, scaling to billions of uses, and many other factors. Google's "study" in particular significantly underestimated actual energy consumption.

Zeke Hausfather has now calculated the magnitude of this gap in an analysis on The Climate Brink. A major blind spot remains: no one outside AI labs knows the actual energy cost per token of a cutting-edge model. Keep in mind that Hausfather's figures are reasonable estimates but remain estimates.

They are also based on current usage patterns. AI labs are already working at the scale of agent-based systems that operate autonomously on tasks for days, weeks, or even months. This could lead to an exponential increase in energy consumption.

3.2 Billion Tokens in Eight Weeks

Hausfather tracked his usage in detail over eight weeks. Claude Code stores complete local logs of each session, including the exact token counts reported by the API for each model call.

His 1,138 typed requests triggered over 14,000 model calls, averaging twelve per request. Each request processed an average of 2.9 million tokens. For comparison, a typical chat exchange without reasoning or web searches hovers around 1,000 tokens.

Energy consumption per AI task spans five orders of magnitude, ranging from 0.24 Wh for a Gemini chat request to several kWh for a full day of Claude Code. The estimates published in blue, the values measured by the author in orange.

In total, his Claude Code processed 3.2 billion tokens. Of these, 96% were cache reads, as at each of the 14,000 steps, the agent rereads all its accumulated context. The text that Hausfather actually sees on the screen, the model output, represents only 0.4% of all tokens processed.

His best estimate for total consumption is around 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 electrical values relies on three independent methods with different assumptions.

Per request, this amounts to about 150 Wh, or about 600 times more than a median chat request. "A 'request' is ultimately not a unit of AI usage any more than 'trips' is a measure of driving; it's the distance traveled that matters," writes Hausfather.

A Single Day of Agent Use Consumes More Than Two Refrigerators

Hausfather's median session of Claude Code consumed about 0.6 kWh, fifty times the electricity needed to charge a phone. His average daily use of Claude Code reached 3.0 kWh (range: 1.2 to 5.9 kWh), more than the daily consumption of two refrigerators.

On his most intensive day, when several parallel agents processed a vast geodata analysis, he estimates that 11 kWh were consumed. This is more than a third of the daily electricity consumption of an average American household, according to Hausfather.

A Year of Intensive Agent Use Equals One Clothes Dryer

Scaled over a full year, Hausfather's use of Claude Code would consume about 1.1 MWh of data center electricity (0.4 to 2.2 MWh), roughly one-tenth of what an average American household uses annually. Based on the average U.S. electricity mix, this translates to about 370 kg of CO₂ equivalents per year.

The annual CO₂ emissions from the author's use of Claude Code (368 kg, orange) compared to daily activities. Ten daily chat requests produce only 0.3 kg per year, while intensive use of the agent equals that of an electric clothes dryer.

This represents slightly more than the operation of an electric clothes dryer for a year (262 kg) and about half of a round-trip economy flight from San Francisco to New York (700 kg), although Hausfather notes that the flight figure only covers direct CO₂ emissions. This accounts for about eight percent of the annual emissions of a typical gasoline car in the U.S. and about two percent of the average annual carbon footprint of an American. "It's both a significant source of emissions and a relatively modest part of my total carbon footprint," writes Hausfather.

Clean Electricity is More Important Than Reducing Consumption

Hausfather does not advocate for sacrifice or guilt. The personal restraint of a small group of heavy users "is not going to bend the curves," he says, adding that directing simple tasks to smaller models makes sense, as they consume five to seven times less energy per token than cutting-edge models.

But the biggest lever is the carbon intensity of the electricity itself, he argues. If the same workload operated primarily on clean energy, the carbon footprint would decrease by about 90 percent. The problem is that nearly three-quarters of the on-site energy production expected for U.S. data centers runs on natural gas.

This is where Hausfather sees an opportunity. AI companies bring enormous capital and unusual urgency. If this money is invested in clean energy, grid expansion, and advanced technologies like geothermal or nuclear energy, the AI boom could leave the grid cleaner than it found it.