AI Agents Use 600x More Energy Than Chat Prompts, Study Finds

AI agents could consume dramatically more electricity than conventional chatbot interactions, according to an analysis by climate scientist Zeke Hausfather. His tracking of coding-agent use over eight weeks suggests that agentic AI workloads can require roughly 600 times as much electricity per user prompt as a typical AI chat prompt.

The findings highlight a growing challenge for the AI industry. While individual chatbot prompts can consume relatively small amounts of electricity, AI agents perform multiple model calls, repeatedly process context and use tools to complete tasks. As companies increasingly shift toward autonomous AI systems capable of working for hours or days, their energy requirements could become significantly larger.

AI Agents Consume Far More Energy

Hausfather tracked his personal use of Anthropic’s Claude Code programming agent for eight weeks. During that period, he entered 1,138 prompts that triggered more than 14,000 model calls and processed approximately 3.2 billion tokens.

His estimated electricity consumption for the workload was around 170 kilowatt-hours, with an estimated range of 70 kWh to 330 kWh depending on the calculation method. That translates to approximately 150 watt-hours per prompt.

For comparison, Google has estimated that a median Gemini text prompt consumes around 0.24 watt-hours, while OpenAI CEO Sam Altman previously estimated an average ChatGPT query at around 0.34 watt-hours.

Using those figures as benchmarks, Hausfather’s Claude Code workload consumed roughly 600 times more electricity per prompt than a median chatbot interaction. However, the comparison is an estimate rather than a direct laboratory measurement because the actual energy consumption per token of leading AI models is not publicly known.

Why AI Agents Use So Much More Power

A conventional chatbot interaction can involve a user sending a question and receiving a response from an AI model.

An AI agent works differently.

Instead of producing one response, an agent can break a task into multiple steps, generate plans, call tools, inspect results, modify its approach and repeat the process until the task is completed.

This can result in dozens or even hundreds of model calls for a single user request.

In Hausfather’s Claude Code logs, his 1,138 prompts triggered more than 14,000 model calls, averaging around 12 calls per prompt. Each prompt processed an average of approximately 2.9 million tokens.

The difference illustrates why measuring AI energy use simply by counting “prompts” can be misleading.

Billions of Tokens Are Processed Behind the Scenes

One of the most striking findings from the analysis was the amount of context processed by the agent.

Claude Code processed approximately 3.2 billion tokens during the eight-week period.

About 96% of those tokens were cache reads, according to Hausfather’s analysis. The agent repeatedly re-read accumulated context as it worked through tasks.

The text actually displayed to the user represented only around 0.4% of all processed tokens.

This highlights an important characteristic of agentic AI: the amount of computation happening behind the scenes can be vastly larger than the amount of text a user sees.

A Single Day Can Consume Several Kilowatt-Hours

Hausfather’s analysis also shows how energy consumption can increase during intensive agent workloads.

His median Claude Code session consumed approximately 0.6 kWh of electricity.

On an average day of use, consumption reached about 3.0 kWh, with individual days ranging between 1.2 kWh and 5.9 kWh.

On his most intensive day, when several agents were working in parallel on a large geodata analysis, estimated consumption reached approximately 11 kWh.

That is substantially different from the energy requirements associated with a handful of ordinary text chatbot queries.

Yearly Energy Use Could Reach 1.1 MWh

If Hausfather’s intensive agent usage pattern were maintained for an entire year, the estimated electricity consumption would reach approximately 1.1 megawatt-hours.

The estimated range is around 0.4 MWh to 2.2 MWh depending on the assumptions used.

Based on the average US electricity mix, Hausfather estimates that this level of usage would correspond to approximately 370 kilograms of CO₂-equivalent emissions annually.

That would be comparable to the emissions associated with running an electric clothes dryer for a year, according to the analysis.

The comparison does not mean every AI user will generate this level of emissions. Hausfather’s workload was intensive and focused on coding-agent usage.

Traditional Chatbots Have a Much Smaller Footprint

The contrast between conventional chatbots and autonomous agents is significant.

Hausfather estimates that ten ordinary AI chat prompts per day would produce only around 0.3 kilograms of CO₂ emissions annually under the assumptions used in his analysis.

Heavy agent usage, by contrast, could generate hundreds of kilograms of annual emissions.

This suggests that the environmental impact of AI will depend increasingly on how people use AI rather than simply how many people use it.

A simple question-answering interaction and an autonomous agent working through a complex coding or research project can represent vastly different workloads.

Reasoning Models Add to the Energy Demand

The energy estimates commonly cited for AI prompts can also fail to capture the growing use of reasoning models.

Reasoning systems can generate significantly more internal tokens before producing a final response.

Multimodal AI, image generation, video generation, coding agents and multi-agent systems can also require substantially more computation than a basic text prompt.

The Decoder noted that published figures from Google and OpenAI do not necessarily capture all these newer workloads.

As AI systems become more capable, therefore, a single “query” is becoming a less useful unit for measuring computing demand.

AI Agents Are Designed to Work Autonomously

The biggest potential change could come from agents that operate without continuous human supervision.

Today’s coding agents can already work through multi-step programming tasks.

Future systems are expected to operate for much longer periods, potentially handling research, software development, data analysis, customer service and other business processes with limited human intervention.

AI companies are actively developing systems designed to operate autonomously for extended periods.

If these systems run continuously, electricity demand could increase substantially even if the number of human prompts remains unchanged.

Data Centres Face a New Challenge

The growth of agentic AI has implications for data-centre operators.

Traditional AI inference workloads can often be optimized around relatively predictable request patterns.

Agentic workloads are more fragmented.

An agent may alternate between model inference, database queries, web searches, code execution and other tools.

Recent research into agentic AI infrastructure has found that these workflows repeatedly move between CPU and GPU resources and can create uneven, bursty workloads. This can make conventional data-centre architectures less efficient.

As agent adoption increases, data centres may therefore need new architectures designed specifically for these workloads.

Energy Efficiency Will Become More Important

AI companies have been investing heavily in more efficient models and hardware.

Smaller models can handle simpler tasks while consuming less computing power.

Hardware improvements can also increase the amount of AI computation performed for each unit of electricity.

Software optimization, caching and better scheduling can further reduce energy requirements.

However, efficiency improvements can be offset if users and businesses dramatically increase the amount of AI computation they perform.

This is sometimes described as a rebound effect: cheaper and more efficient computing can encourage greater usage.

Clean Energy Could Reduce Emissions

Hausfather argues that shifting data centres toward cleaner electricity is one of the most important ways to reduce the climate impact of AI.

The same amount of AI computation can produce very different emissions depending on the electricity source powering the data centre.

A workload supplied primarily by low-carbon electricity would have a significantly smaller carbon footprint than the same workload powered largely by fossil fuels.

This means AI companies’ choices around renewable energy, nuclear power and other low-carbon electricity sources could become increasingly important as agent workloads expand.

AI’s Environmental Impact Is Not Just About Electricity

Electricity consumption is only one component of AI’s environmental footprint.

Large AI data centres also require:

• Advanced GPUs and other chips
• Servers
• Cooling systems
• Water
• Networking equipment
• Buildings and electrical infrastructure
• Semiconductor manufacturing

The production of AI hardware itself requires significant resources.

As companies deploy increasingly powerful AI infrastructure, the environmental impact therefore extends beyond the electricity used during model inference.

The Economics of AI Agents

Energy consumption also has a direct economic dimension.

More computation means higher infrastructure costs.

If an agent makes dozens of model calls to complete a task, the provider must pay for additional GPU time, networking, storage and supporting infrastructure.

This could influence the pricing of agentic AI products.

AI companies may increasingly charge users based on usage, task complexity or compute consumption rather than offering unlimited access at a flat subscription price.

The economics could become particularly important for autonomous agents that operate continuously.

Companies May Need Compute Budgets

Businesses deploying AI agents at scale could eventually introduce internal compute budgets.

Instead of allowing an AI agent to run indefinitely, organizations could establish limits around:

• Maximum task duration
• Number of model calls
• Token consumption
• Tool usage
• GPU resources
• Energy consumption
• Cost per completed task

Such controls could help businesses balance automation benefits against infrastructure expenses.

More Computation Does Not Always Mean Better Results

Another important issue is efficiency.

An agent may make many additional model calls without necessarily improving the final result proportionally.

Recent research into coding-agent workloads found that prompt formulation and agent-harness design can significantly affect reasoning costs. In some experiments, certain prompting patterns increased reasoning-token consumption by several times without improving task correctness.

This suggests that AI developers have another opportunity to reduce energy consumption: making agents more efficient rather than simply making them more powerful.

Better Agent Design Could Reduce Waste

AI agents can potentially become more efficient through better planning and orchestration.

Instead of repeatedly processing the same context, systems could maintain more efficient memory structures.

Agents could also determine when a task genuinely requires a powerful reasoning model and when a smaller model would be sufficient.

Other approaches include:

• Better caching
• Smaller specialized models
• More efficient context management
• Fewer unnecessary tool calls
• Better task planning
• Early stopping
• Improved model routing
• More efficient hardware

These changes could reduce the amount of computation required to accomplish the same task.

What This Means for AI Users

The findings do not mean people should stop using AI.

Instead, they show that different AI tasks have very different computing requirements.

A simple factual question is fundamentally different from asking an autonomous agent to spend hours writing and testing software.

Users can reduce unnecessary computation by giving clearer instructions, avoiding repeated failed attempts and choosing simpler models when advanced reasoning is not necessary.

For organizations, the focus will increasingly be on measuring the cost and efficiency of AI workflows rather than simply counting the number of AI interactions.

The Biggest Issue Is Scale

The environmental impact becomes more important when millions of users adopt agents.

One intensive agent session may consume a manageable amount of electricity.

But if millions of people run autonomous agents simultaneously, the combined demand could become substantial.

The AI industry is already investing billions of dollars in data-centre capacity to support growing demand.

The expansion of agentic AI could accelerate that investment cycle.

AI Infrastructure Could Become a Major Electricity Consumer

As AI models become more capable and agents become more autonomous, electricity availability could become an increasingly important constraint.

Data-centre operators are already looking for additional power capacity in regions where AI infrastructure is expanding.

The issue is no longer simply whether companies can buy enough GPUs.

They also need sufficient electricity, cooling capacity, grid connections and physical infrastructure to operate those GPUs.

Agentic AI could increase the pressure on all of these resources.

Industry Impact

The findings highlight a major shift in the economics and environmental footprint of artificial intelligence.

The early AI era was dominated by simple chatbot interactions, where individual requests consumed relatively small amounts of electricity.

The next phase is increasingly focused on agents that reason, use tools and perform multiple steps autonomously.

That transition could dramatically increase computing demand.

For AI companies, improving model efficiency and securing low-carbon electricity will become increasingly important. For businesses, understanding the cost of agentic workloads will be critical before deploying autonomous systems at scale.

Looking Ahead

The estimated 600-fold difference between a typical chatbot prompt and an intensive AI-agent prompt illustrates how dramatically computing requirements can change when AI moves from answering questions to performing multi-step tasks. Hausfather’s eight-week Claude Code analysis estimated roughly 170 kWh of data-centre electricity for 1,138 prompts, with more than 14,000 model calls and 3.2 billion tokens processed.

The figures remain estimates rather than precise measurements of every AI agent, and actual energy consumption varies significantly by model, hardware, task and data-centre efficiency. Nevertheless, the analysis highlights an increasingly important issue for the AI industry: as agents become capable of working autonomously for longer periods, the biggest challenge may no longer be simply making AI more intelligent, but making that intelligence affordable and energy-efficient at scale.