AI Agents Consume Roughly 600 Times More Energy Than a Chat Prompt, New Analysis Finds

Google and OpenAI have spent the past year reassuring the public that an AI query barely uses more electricity than a short TV clip. A new real-world audit from climate scientist Zeke Hausfather suggests the truth for autonomous AI agents is far more sobering: they can consume roughly 600 times more energy per prompt than a standard chat exchange. The findings, published August 8, 2026, in an analysis on The Climate Brink, add fresh urgency to questions about AI's environmental footprint as the industry pivots toward always-on autonomous agents. For ongoing reporting on the sustainability debate shaping the industry, follow our AI industry coverage.

A Reality Check on Big Tech's Energy Claims

Last year, Google reported that a median Gemini text prompt uses just 0.24 watt-hours (Wh) of electricity — less than nine seconds of television. OpenAI CEO Sam Altman offered a similar figure, estimating the average ChatGPT query at roughly 0.34 Wh, on par with a Google search from 2009.

Both numbers describe a narrow, best-case scenario: a single text exchange with no extended reasoning, no web searches, no multimodal processing, and no looping agent behavior. As Hausfather points out, they do not account for reasoning models that generate many more tokens per query, multi-agent systems that run autonomously, or the cascading context that agents re-read at every step.

Eight Weeks, 3.2 Billion Tokens

Hausfather, a climate scientist, tracked his own usage of Anthropic's Claude Code coding agent in meticulous detail over eight weeks. Because Claude Code stores complete local logs of every session, including the exact token counts reported by the API for each model call, he could reconstruct a precise picture of what an agent actually demands.

The numbers are striking. His 1,138 typed prompts triggered more than 14,000 model calls — an average of about twelve per prompt. Each prompt processed an average of 2.9 million tokens, compared with roughly a thousand tokens for a typical chat exchange. In total, his agent processed 3.2 billion tokens over the period.

A striking detail: roughly 96 percent of those tokens were cache reads. At each of the 14,000 steps, the agent re-reads its entire accumulated context before deciding what to do next. The text Hausfather actually sees on screen — the model's output — accounts for just 0.4 percent of all the tokens processed. The hidden computational cost happens almost entirely behind the scenes.

Translating Tokens Into Electricity

Because no outside party knows the exact energy cost per token for a frontier model, Hausfather converted his token counts to electricity values using three independent estimation methods with different assumptions. His best estimate puts total consumption at about 170 kilowatt-hours (kWh) of data center electricity over eight weeks, with an uncertainty range of 70 to 330 kWh.

Per prompt, that works out to roughly 150 Wh — about 600 times the energy of a median chat prompt. His median Claude Code session consumed about 0.6 kWh, roughly fifty times the electricity needed to charge a smartphone. On an average day, his agent usage hit 3.0 kWh (ranging from 1.2 to 5.9 kWh), more than the daily draw of two household refrigerators.

His most intensive day — when several parallel agents ran an extensive geodata analysis — consumed an estimated 11 kWh, more than a third of the daily electricity use of an average U.S. household.

What It Means for a Year of Agent Use

Scaled to a full year, Hausfather's agent-based workflow would consume roughly 1.1 megawatt-hours (MWh) of data center electricity, with a range of 0.4 to 2.2 MWh — about one-tenth of what an average U.S. household uses annually. Based on the average U.S. electricity mix, that translates to approximately 370 kilograms of CO2-equivalent emissions per year.

For context, that is more than running an electric clothes dryer for a year (about 262 kg) and roughly half the direct emissions of a round-trip economy flight from San Francisco to New York (about 700 kg). Hausfather notes it amounts to about eight percent of the annual emissions of a typical U.S. gas-powered car, and roughly two percent of the average American's yearly carbon footprint.

"This is simultaneously a large emissions source and a relatively modest part of my total carbon footprint," he wrote.

The Real Levers: Clean Power and Smarter Routing

Hausfather is careful not to frame the issue around personal guilt. Restraint by the small group of heavy users "is not going to bend any curves," he argues. Instead, he highlights two structural levers.

First, routing simpler tasks to smaller models makes a meaningful difference. Smaller models use roughly five to seven times less energy per token than frontier models. Second — and most importantly — the carbon intensity of the electricity itself is the dominant factor. If the same workload ran on largely clean power, the carbon footprint would fall by roughly 90 percent.

The problem is the fuel behind the boom. According to Hausfather, nearly three-quarters of planned on-site power generation for U.S. data centers runs on natural gas, meaning the rapid expansion of AI computing is being underwritten in large part by fossil fuels.

A Warning as Agents Scale

Hausfather's estimates, while reasonable, remain estimates — and they reflect current usage patterns. AI labs are already racing to deploy agent-based systems that run autonomously for days, weeks, or even months. If that vision materializes, energy consumption per "user" could climb far higher still.

His central message is that the per-query figures Google and OpenAI have publicized distort the reality of how the technology is evolving. "A 'prompt' is ultimately not a unit of AI use any more than 'trips' is a measurement of driving; it's how far you go that matters," Hausfather wrote.

Stay Ahead of AI

As autonomous agents move from novelty to default infrastructure, their hidden energy bill will become harder to ignore. For continuous, fact-checked reporting on AI research, sustainability, and the companies shaping the field, bookmark our homepage and read more AI news → at https://aibuzzwire.news.