A new study has produced a scary headline: AI agents can use up to 136.5 times more electricity than a standard chatbot for a single query. The number went viral. According to the research from KAIST, reported by the Korea Times, Forbes, Gizmodo and Digital Trends, it is real - but it is also widely misunderstood. Here is what it actually means. For the basics, see our what is an AI agent guide.
What the study found
A team at KAIST (the Korea Advanced Institute of Science and Technology) published a paper around 5 July 2026 on the cost of "dynamic reasoning" - the loop-and-call style of modern AI coding agents. They measured the operational power of chatbot-style answers against agent-style work.
The peak result: an agent that browses the web, writes code and chains together multi-step plans can draw up to 136.5 times more energy per question than a plain generative-AI reply.
The number is a peak, not an average
This is where the viral headline misleads. 136.5x is the worst case, not the typical case. It describes heavy, tool-calling, multi-step agents - not every agent on every task. A simple agent doing a short job is nowhere near that figure.
Reporting from Forbes made the same point: the viral number is a peak, and treating it as an average overstates the everyday cost. The honest reading is that agents can be far more expensive, not that they always are.
The real problem: idle GPUs
The more useful finding is why agents cost so much. An agent does not answer in one shot. It plans, calls a tool, waits, reads the result, then calls the model again - often many loops per task.
While the agent waits on an external tool or a slow web page, the expensive GPU is switched on but not computing. The study found GPUs can sit idle up to 54.5% of the time. You are paying full power for hardware that is, for much of the task, just waiting. That idle time - not the raw model size - is the hidden energy tax.
What it means for you
Two practical takeaways:
- Match the tool to the task. A single, well-formed model call is far cheaper than an agent that loops. For a simple question, do not spin up a multi-step agent.
- Value the loops. When an agent's browsing, coding and checking genuinely saves you work, the extra energy can be worth it. The goal is efficiency, not avoidance.
The honest takeaway
Two caveats keep this straight. First, 136.5x is a peak from one study, reported widely but still a single research group's measurement - useful, not gospel. Second, the point is not that agents are bad; it is that their power cost scales with how much they loop and wait.
The honest read: AI agents are a real step up in capability and a real step up in energy use, driven mostly by idle hardware during multi-step reasoning. Use them where the multi-step work pays off, and reach for a plain model call when it does not. To choose efficient models in the first place, our best coding LLMs 2026 overview helps.
