AI agents use 600 times energy in one Claude Code analysis
A climate scientist’s eight-week Claude Code log shows why a simple prompt is a poor proxy for agentic AI power use.
By Colin Brandt · Enterprise Reporter
· 3 min read
AI agents use 600 times energy is the headline result from climate scientist Zeke Hausfather’s analysis of his own Claude Code activity, published Aug. 5. The figure describes one intensive eight-week coding-agent workload, not a standard electricity cost for every agent, task or provider.
Hausfather reported that 1,138 prompts he typed into Claude Code triggered more than 14,000 model calls and processed 3.2 billion tokens. He estimated the work consumed 170 kilowatt-hours of data-center electricity, or about 150 watt-hours per typed prompt.
That estimate is about 625 times Google’s reported 0.24 Wh median Gemini text prompt, which Hausfather used as the comparison point. He characterized the result as roughly 600 times, while his stated 60 Wh to 290 Wh per-prompt uncertainty range works out to about 250 to 1,200 times the Gemini benchmark.
Why do AI agents use more energy than chat prompts?
An agent can turn one human instruction into a sequence of model requests: planning work, writing or reviewing code, running commands, reading results and trying again. In Hausfather’s logs, each typed prompt generated about 12 model calls and processed an average of 2.9 million tokens. He contrasted that with roughly 1,000 tokens for a typical web chat exchange without reasoning or web search.
The bulk of the compute in this case was not visible in the final answer. Hausfather said 96% of the 3.2 billion tokens were cache reads, as the agent repeatedly reread accumulated working context. That dynamic makes a count of user prompts a weak measure of resource use for long-running, iterative workflows.
What was logged and what was estimated?
Logged: Claude Code’s local transcripts supplied API-reported token counts for each model call, according to Hausfather.
Estimated: The conversion from tokens to electricity used energy factors and assumptions, including assumptions about the energy cost of cache reads. Hausfather put the eight-week total between 70 kWh and 330 kWh across methods and assumptions.
The distinction limits what can be concluded from the analysis. It documents a detailed example of heavy use by one person, rather than a direct meter reading of model energy use or a universal agent benchmark. Hausfather described himself as a Stripe employee who uses current AI tools more than most people.
His results nevertheless sit within a broader range cited in the analysis. A Watershed framework places agentic workflows that make five to 50 frontier-model calls at 50 Wh to 500 Wh, while published estimates across AI tasks span more than five orders of magnitude. Product teams comparing agent architectures therefore need to track model calls, token volume and context handling alongside per-user request counts.
Hausfather annualized his observed pattern at 1.1 MWh of data-center electricity, with a 0.4 MWh to 2.2 MWh range. That is an extrapolation, not a forecast. Carbon emissions would also vary with the electricity mix, so the 600-times comparison is about electricity consumed per interaction rather than a fixed emissions figure.
This story draws on original reporting from The Decoder.
