Simple AI chat prompts can look cheap from an energy perspective. Agentic AI is a different kind of workload. An analysis by climate scientist Zeke Hausfather, based on his own Claude Code logs, shows how a single typed instruction can expand into many model calls, billions of processed tokens, and a much larger electricity footprint.
Why one AI prompt no longer means one model response
Google and OpenAI have pointed to low energy figures for ordinary AI queries. Google says a median Gemini text prompt uses 0.24 watt-hours (Wh), while OpenAI CEO Sam Altman estimated the average ChatGPT query at 0.34 Wh. Those numbers may describe a simple exchange, but they do not capture the way many people now use AI systems.
AI agents often do more than answer a question. They may reason through a task, call models repeatedly, review accumulated context, generate code, process different kinds of inputs, or coordinate work across multiple steps. That makes the word “prompt” a weak unit for measuring real AI use.
Hausfather puts the issue plainly:
"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,"
The point is not that every AI request has the same footprint. It is that agentic systems can turn one user input into a long chain of computation that is mostly invisible to the person typing.
What the Claude Code logs showed
Hausfather tracked eight weeks of his own Claude Code activity. The tool stores local logs for each session, including token counts reported by the API for each model call. That allowed him to measure the workload behind his visible interactions.
Across that period, 1,138 typed prompts triggered over 14,000 model calls. That averages twelve model calls per prompt. Each prompt processed an average of 2.9 million tokens, while a typical chat exchange without reasoning or web searches runs about a thousand tokens.
In total, Claude Code processed 3.2 billion tokens during the eight-week period. The overwhelming majority of that work was not the output Hausfather read on screen. Of all processed tokens, 96 percent were cache reads, because the agent re-reads its accumulated context at each of the 14,000 steps. The visible output accounted for just 0.4 percent of the total.
That distinction matters for AI energy use. A user may experience a task as a small number of instructions and responses, while the underlying system repeatedly processes a much larger body of context. In agent workflows, the costly part may be the repeated background work, not the final text that appears in the interface.
The estimated electricity cost was far above simple chat
Hausfather’s best estimate for the eight weeks of Claude Code use was about 170 kWh of data center electricity. The uncertainty range was 70 to 330 kWh. He directly measured token counts from the Claude Code logs, while the conversion from tokens to electricity relied on three independent methods with different assumptions.
On a per-prompt basis, the estimate comes to roughly 150 Wh. That is about 600 times as much as a median chat prompt. The comparison helps explain why low figures for simple AI prompts can be misleading when applied to agent-based coding or analysis workflows.
The session-level and day-level numbers are also revealing. Hausfather’s median Claude Code session consumed about 0.6 kWh, which is fifty times the electricity needed to charge a phone. His average day using Claude Code reached 3.0 kWh, with a range of 1.2 to 5.9 kWh. That was more than the daily draw of two refrigerators.
On his most intensive day, several parallel agents were working through an extensive geodata analysis. Hausfather estimated that day consumed 11 kWh. According to Hausfather, that is more than a third of the daily electricity consumption of an average U.S. household.
Annual emissions depend heavily on the power source
If that level of agent-based Claude Code use continued for a full year, Hausfather estimates it would consume about 1.1 MWh of data center electricity, with a range of 0.4 to 2.2 MWh. That is roughly one-tenth of what an average U.S. household uses annually.
Using the average U.S. electricity mix, the annual footprint would be about 370 kg of CO₂ equivalents. The source compares that with slightly more than running an electric clothes dryer for a year, listed at 262 kg, and about half of a round-trip economy flight from San Francisco to New York, listed at 700 kg. Hausfather notes that the flight figure covers only direct CO₂ emissions.
The same annual agent use would equal 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. Hausfather describes the scale this way:
"This is simultaneously a large emissions source and a relatively modest part of my total carbon footprint,"
Cleaner data center power is the main lever
Hausfather does not frame the answer mainly as personal restraint. He says restraint by the small group of heavy users "is not going to bend any curves." Still, the analysis notes that routing simple tasks to smaller models makes sense, because they use five to seven times less energy per token than frontier models.
The larger issue is the carbon intensity of the electricity used to run the workload. If the same work ran on largely clean power, Hausfather argues, the carbon footprint would fall by about 90 percent.
That is where the future of AI infrastructure becomes important. The source notes that nearly three-quarters of planned on-site power generation for U.S. data centers runs on natural gas. At the same time, AI companies have enormous capital and unusual urgency. If that investment goes into clean energy, grid expansion, and advanced technologies like geothermal or nuclear power, Hausfather argues the AI boom could help leave the grid cleaner than it found it.
The takeaway is straightforward: AI agents are not just chatbots with longer answers. They are multi-step systems that can amplify computation behind a single user instruction. Measuring their impact requires looking at tokens, model calls, data center electricity, and the power mix that supplies it.