Anthropic's Claude Code uses 600 times more power than chat

Climate scientist Zeke Hausfather tracked his usage of Anthropic's programming agent, Claude Code, over an eight-week period to measure its actual energy footprint. His analysis, published on The Climate Brink, revealed that 1,138 typed prompts triggered more than 14,000 model calls, averaging 12 calls per prompt. This activity processed 3.2 billion tokens, with 96 percent consisting of cache reads as the agent re-read its context at each step, while final outputs accounted for just 0.4 percent. In total, the agent consumed approximately 170 kilowatt-hours of electricity, translating to roughly 150 watt-hours per prompt.

This consumption is roughly 600 times higher than the figures typically cited by major AI labs for simple chat queries. For instance, Google reports that a median Gemini text prompt uses just 0.24 watt-hours, while OpenAI CEO Sam Altman estimated a ChatGPT query at 0.34 watt-hours. In contrast, Hausfather's median Claude Code session consumed 0.6 kilowatt-hours, and his average daily usage reached 3.0 kilowatt-hours, with a peak day of 11 kilowatt-hours during a complex geodata analysis. Annually, this level of agent use would consume 1.1 megawatt-hours and generate 370 kilograms of carbon dioxide equivalents, comparable to running an electric clothes dryer.

For developers and practitioners, these findings demonstrate that agentic workflows carry a vastly different resource profile than standard LLM interactions. To mitigate this impact, practitioners can route simpler tasks to smaller models, which consume five to seven times less energy per token than frontier models. Ultimately, however, Hausfather argues that individual restraint will not shift the curve. Instead, the AI industry must leverage its massive capital to fund clean energy grid expansions, nuclear power, and geothermal projects to offset the carbon intensity of data centers.