A detailed analysis of Claude Code usage suggests that AI agents can consume dramatically more electricity than the simple chat prompts often used to describe AI’s energy footprint. Climate scientist Zeke Hausfather tracked his own programming-agent activity for eight weeks and estimated roughly 170 kilowatt-hours of data center electricity for 1,138 typed inputs.

That works out to about 150 watt-hours per input—approximately 600 times the 0.24 to 0.34 watt-hours cited for a typical text query by Google and OpenAI. The comparison is not a universal measurement of every agent workload, but it highlights how much energy can be hidden behind a single user request when software agents make repeated model calls, process large contexts, and work autonomously.

What Hausfather measured in Claude Code

Hausfather reported his findings through The Climate Brink, with coverage from The Decoder. His local Claude Code logs recorded token counts and model calls across eight weeks of use. The 1,138 prompts triggered more than 14,000 model calls—about 12 calls per input—and processed approximately 3.2 billion tokens.

The largest share of that processing came from context management rather than text displayed to the user. About 96 percent of the tokens were cache reads, according to the analysis, because the agent repeatedly re-read its accumulated context at successive steps. Model output visible on screen represented only about 0.4 percent of the processed tokens.

Hausfather estimated total consumption at 170 kWh, but gave a wide uncertainty range of 70 to 330 kWh. The token counts came from the Claude Code logs; converting those counts into electricity use required three separate estimation methods and assumptions about energy consumed per token.

The resulting workload was uneven. A median session used about 0.6 kWh, while average daily consumption reached 3.0 kWh, with a range from 1.2 to 5.9 kWh. The most intensive day, involving several parallel agents working on a geodata analysis, was estimated at 11 kWh.

Why simple prompt comparisons fall short

Google has said that a median Gemini text prompt uses 0.24 watt-hours. OpenAI CEO Sam Altman has estimated an average ChatGPT query at 0.34 watt-hours. Those figures describe narrow query types and are not necessarily comparable with a coding agent that reasons through a task over many steps.

The distinction matters because “one prompt” is a user-interface event, not a consistent unit of computation. A conventional chat exchange may involve roughly a thousand tokens, according to Hausfather’s comparison. His Claude Code inputs averaged about 2.9 million processed tokens, including repeated context and intermediate calls.

Reasoning models, web searches, image and audio processing, code generation, and multi-agent workflows can all change the amount of computation behind a request. An agent may also retry operations, inspect files, run tests, revise code, and call tools before returning a result. A product dashboard that reports only the initial prompt could therefore obscure the actual workload.

This does not establish that every use of AI agents consumes 600 times as much electricity as a chat request. The figure is an estimate from one intensive user’s Claude Code pattern, compared with published low-end query estimates. It does show why average prompt figures should not be treated as a complete accounting of agent-based AI.

Evidence, uncertainty, and the carbon question

The analysis provides unusually concrete usage data but does not directly measure electricity at the data center. Hausfather’s uncertainty range is substantial, and the energy-per-token assumptions are not independently verifiable from outside AI companies. The Decoder also notes that outsiders do not know the precise energy cost per token for leading commercial models.

The carbon impact depends on more than electricity volume. Hausfather projected that maintaining this level of use for a year would require about 1.1 megawatt-hours of data center electricity, with a range of 0.4 to 2.2 MWh. Using the average U.S. electricity mix, he estimated roughly 370 kilograms of carbon dioxide equivalent annually.

That figure would be lower on a mostly clean grid and higher where power comes from carbon-intensive generation. Hausfather estimates that running the same workload on largely clean electricity could reduce its carbon footprint by about 90 percent. His conclusion is that the electricity source may be a more important lever than asking a small number of heavy users to stop using agents.

He also argues that smaller models are useful for routine tasks because they consume about five to seven times less energy per token than frontier models, according to his analysis. That is a directional recommendation rather than a universal purchasing rule: model efficiency, quality, latency, hardware, and the number of retries all affect the final result.

What the findings mean for AI builders and buyers

For developers, the immediate lesson is to measure complete agent runs rather than count user prompts. Teams should track model calls, input and output tokens, cache reads, tool invocations, retries, parallel workers, and the total time spent on a task. These metrics can reveal that an apparently inexpensive feature is generating substantial backend work.

The findings also make routing strategies more consequential. A smaller model could handle classification, file lookup, formatting, or simple code changes, while a more capable model is reserved for tasks that genuinely require deeper reasoning. Context pruning, retrieval design, caching, and limits on autonomous loops may reduce both cost and electricity use, although aggressive limits can also reduce reliability.

Enterprise buyers should ask vendors how they calculate energy and carbon claims. A per-query number may exclude hidden reasoning, cached context, tool calls, or data-center overhead. Buyers evaluating coding assistant or workflow-agent deployments need workload-specific estimates, not only averages from lightweight chat traffic.

The analysis also raises a deployment question for products designed to run for days or weeks. Longer autonomous tasks could multiply energy demand well beyond the usage pattern Hausfather recorded. That possibility is not a forecast, but it is a reason to include energy, cost, and termination controls in agent architecture before scaling usage across an organization.

What to watch next

The next useful signal will be independent measurement of energy per token across model families, hardware configurations, and agent frameworks. Clearer reporting from AI labs could make vendor claims more comparable, particularly if companies disclose whether figures include reasoning, caching, tool calls, and cooling overhead.

Product teams should also watch for practical controls such as model routers, context-budget settings, agent step limits, and workload-level energy dashboards. On the infrastructure side, the share of new data-center power supplied by natural gas, alongside investment in clean energy, grid expansion, geothermal power, or nuclear generation, will determine how much of the additional electricity translates into emissions.

Creati.ai perspective

Hausfather’s analysis does not show that AI agents are inherently wasteful, nor does it invalidate lower energy estimates for simple chat. It shows that the unit being measured matters. For an agent, the user’s prompt may be only the beginning of a long computational workflow.

For builders and enterprise buyers, the practical response is better accounting: measure the full task, route work to the smallest capable model, and design limits around autonomous execution. The energy debate will become more useful when product-level reporting reflects what agents actually do rather than what users type.