Anthropic’s Claude Code burned roughly 170 kWh of electricity during an eight‑week test, a figure that translates to about 150 Wh per prompt – an order of magnitude higher than the 0.34 Wh typical of a ChatGPT request, according to Sam Altman’s earlier comment.
Measurement methodology and raw numbers
The German tech news site Heise reported that climate researcher Zeke Hausfather, who works at Stripe and Berkeley Earth, logged his own usage of Claude Code from September 2026 through mid‑October 2026. Over that period he entered 1,138 prompts, which triggered more than 14,000 model calls and generated 3.2 billion tokens. Hausfather’s electricity meter recorded an estimated consumption of ~170 kWh, with a tolerance band of 70 kWh to 330 kWh depending on the calculation method used.
All three figures – the 170 kWh total, the 70‑330 kWh range, and the 150 Wh per‑prompt average – are directly quoted from the Heise article “Klimafaktor KI: Warum Agenten wie Claude Code echte Stromfresser sind”.
Per‑prompt energy: Claude Code versus a standard chatbot
Dividing the 170 kWh total by the 1,138 prompts yields an average of about 150 Wh per prompt. By contrast, Sam Altman has previously said a normal ChatGPT query consumes roughly 0.34 Wh. The disparity is illustrated in the table below.
| Metric | Claude Code (Hausfather) | Typical ChatGPT prompt |
|---|---|---|
| Energy per prompt | 150 Wh | 0.34 Wh (quoted by Sam Altman) |
| Source: Heise – Klimafaktor KI | ||
The table makes clear that an autonomous agent such as Claude Code can consume more than 400 times the electricity of a conventional conversational query.
Carbon footprint and household‑scale comparison
Heise also notes that the measured electricity use would emit roughly 370 kg of CO₂. That amount exceeds the annual emissions of a typical clothes dryer (about 262 kg) and falls just short of the yearly emissions of an average electric vehicle, according to the same source.
If Hausfather’s eight‑week consumption rate were sustained for a full year, the total would be about 1.1 MWh (170 kWh × 52 weeks ÷ 8 weeks). This extrapolation is a simple arithmetic extension of the reported figure and is not presented as a quoted claim.
For context, the U.S. Energy Information Administration estimates the average U.S. household uses roughly 10 MWh of electricity per year. Thus, a single power user of Claude Code would consume roughly one‑tenth of a typical household’s annual electricity demand.
The measurement underscores a growing concern among climate analysts: as AI agents acquire more autonomous capabilities – scheduling meetings, drafting contracts, or even completing e‑commerce transactions – their computational intensity rises sharply. The Heise piece frames this as a “real‑world electricity‑eater” problem, warning that the surge in data‑centre construction for AI could outpace current sustainability measures.
Enterprises that plan to embed agents like Claude Code into internal workflows should factor the higher per‑prompt energy cost into their total‑cost‑of‑ownership calculations. For a company that runs thousands of prompts daily, the cumulative electricity bill and associated carbon emissions could become material ESG (environmental, social, governance) considerations.
Regulators may also look to these early measurements when shaping reporting standards for AI‑related energy use. The European Commission’s upcoming AI‑act, for example, could eventually require firms to disclose the electricity consumption of high‑impact models, much as they already report data‑centre PUE (Power Usage Effectiveness).
Outlook and unanswered questions
Hausfather’s study is the first publicly disclosed eight‑week measurement of an AI agent’s power draw, but many variables remain unknown. The Heise article does not break down how much of the electricity was attributable to model inference versus supporting infrastructure such as networking or storage. It also does not disclose the exact hardware configuration of the test environment, which can influence energy efficiency dramatically.
Future research could compare Claude Code’s consumption on different cloud providers, or assess how optimisation techniques – model quantisation, sparsity, or on‑device inference – might narrow the gap with traditional chatbots. Until then, the 150 Wh per‑prompt figure serves as a cautionary benchmark for organisations weighing the productivity gains of autonomous agents against their climate impact.
For readers interested in a broader view of AI‑energy consumption, City AM previously covered the topic in “AI‑energy consumption analysis”.