Climate scientist Zeke Hausfather tracked his Claude Code usage over eight weeks. Result: 1,138 prompts generated over 14,000 model calls and 3.2 billion tokens, for an estimated electricity consumption of 170 kWh from data center operations. Each prompt costs roughly 150 Wh, 600 times more than a simple chat.

The difference lies in how the agent works. At each step, Claude Code rereads the entire accumulated context: 96% of processed tokens are cache reads. The text you see on screen is 0.4% of the total. It’s the same pattern we covered when agentic memory cost became a design criterion: the agent is a loop that goes through the same data thousands of times.

For those deploying agents, energy consumption is an operational metric that hits the bottom line. An intense day of Claude Code consumes as much as two refrigerators; Hausfather’s worst day hit 11 kWh, a third of the daily electricity consumption of an average American home. Projected over a year, heavy agent use produces roughly 370 kg of CO₂ equivalents, slightly more than an annual electric dryer.

The figures Google and OpenAI publish about energy costs “per query” describe chat, not agents. When work shifts from a single question to an autonomous loop running for hours, consumption changes by orders of magnitude. Companies that have deployed agents at scale are discovering this on their own accounts: non-engineers especially burn tokens converting PDFs to markdown without realizing it.

In detail

Public figures on AI energy consumption so far painted a reassuring picture. Google stated that a median text prompt on Gemini consumes 0.24 Wh, less than nine seconds of television. Sam Altman estimated an average ChatGPT query at 0.34 Wh, comparable to a 2009 Google search. Both figures are from 2025 and describe a simple interaction: one question, one answer, no extended reasoning, no web search, no autonomous loop.

Hausfather measured something different. Claude Code is a programming agent that works in loops: it receives an objective, plans a series of steps, calls the model at each step, reads the results, decides what to do next. Over eight weeks, 1,138 prompts triggered 14,000 model calls, an average of twelve per prompt. Each prompt processed an average of 2.9 million tokens, against ~1,000 for a chat without reasoning.

The most significant detail is token composition. 96% are cache reads: at each of the 14,000 steps, the agent rereads the entire accumulated context up to that point. The text the user sees as a response, the model output, is 0.4% of the total. Nearly all the energy goes into reviewing data already seen, not generating new data.

Converting tokens to kilowatt-hours is not straightforward, and Hausfather is explicit about limitations. No one outside the labs knows the actual per-token consumption of a frontier model. Estimates are based on three independent methods with different assumptions, and the uncertainty range is wide: from 70 to 330 kWh over eight weeks, against a point estimate of 170 kWh. The figure with highest confidence is the token count, because Claude Code logs locally with exact API counts.

Projected over a year of intensive use, consumption reaches roughly 1.1 MWh (range: 0.4-2.2 MWh), one-tenth of the annual electricity consumption of an average American home. In terms of emissions, based on the US electricity mix, we’re talking about 370 kg of CO₂ equivalents per year. For comparison: an annual electric dryer produces about 262 kg, a round-trip San Francisco–New York economy flight about 700 kg (direct emissions only).

A parallel thread emerges from the economic side. Companies that have deployed agents at scale find themselves managing token consumption they didn’t anticipate. At Accenture, internal data shows non-engineers drive token consumption, and one of the most costly practices is PDF-to-markdown conversion: documents are converted to images then to text, bloating the context at each step.

The implications for those building with agents are concrete. First, the “prompt” stops being a useful unit of measurement: what matters is how much context the agent processes, how many calls it makes, how many times it rereads. Second, optimization concerns the flow architecture, not just the model or single prompt: compress context, manage cache, avoid passing raw PDFs when a structured summary will do. Third, today’s numbers are conservative: labs are already building agents that run for days or weeks, and consumption could grow exponentially.

Hausfather points to shifting data centers toward clean energy sources as the most important lever for emissions. For those using agents in their work, the immediate lever is different: understand where tokens go, and design flows so they don’t burn rereading things already known.