How much energy does ChatGPT use? A single query uses about 0.34 watt-hours of electricity, according to OpenAI CEO Sam Altman, roughly what an oven draws in a little over one second. Independent analysis by Epoch AI puts a typical GPT-4o query near 0.3 Wh, about the same as one Google search. The catch is that heavier models reverse this; researchers estimate GPT-5 averages around 18 watt-hours per query, with complex answers reaching 40 Wh. Across 2.5 billion queries a day, even the small number adds up to roughly 310 GWh a year.
This guide gives you a dated, sourced answer for every version of the question, so per query, per prompt, per day, and per year. You will see why estimates differ so widely, why the old “ChatGPT uses ten times more power than Google” claim fell apart, and why newer reasoning models are hungrier rather than leaner. We lead with the data, not the panic, and we flag exactly what each number does and does not count.
The Key Takeaways
- OpenAI’s official figure: 0.34 Wh per query. Sam Altman states an average ChatGPT query uses about 0.34 watt-hours, comparable to an oven running for just over a second.
- Independent estimate: ~0.3 Wh for GPT-4o. Epoch AI’s 2025 analysis lands near 0.3 Wh, roughly 10x lower than the old, widely repeated 3 Wh figure.
- GPT-5 is far hungrier. University of Rhode Island researchers estimate a GPT-5 query averages 18.35 Wh and can hit 40 Wh, versus 2.12 Wh for GPT-4, about 8.6x more.
- It is roughly a Google search now, not 10x. The famous “10 times a Google search” claim came from a 2023 guess; modern per-query energy is close to Google’s own 0.3 Wh figure.
- Per query it is tiny; at scale it is not. At 2.5 billion queries a day, ChatGPT draws roughly 850 MWh daily and close to 310 GWh a year, about the electricity of 29,000 US homes.
How Much Energy Does ChatGPT Use Per Query?
A single ChatGPT query uses about 0.34 watt-hours of electricity, the figure OpenAI’s Sam Altman gave in his 2025 essay “The Gentle Singularity.” He framed it as roughly what an oven uses in a little over one second, or a high-efficiency lightbulb in a couple of minutes. An independent 2025 analysis by Epoch AI reached a similar place from the other direction, estimating a typical GPT-4o query at about 0.3 Wh. Both numbers describe an ordinary text answer, not a long research request or a heavy reasoning task.
The reason you see a range instead of one tidy number is that each estimate draws the boundary in a different place. Altman’s figure is OpenAI’s own and is not peer reviewed, it does not define what an “average” query is, and it excludes the energy used to train the model. Epoch AI built its number from hardware math, one second of H100 GPU time per query at 1,500 watts with a 70% utilization factor, which works out to about 1,050 watt-seconds, or 0.3 Wh. The heaviest case is the newest one, where University of Rhode Island researchers estimate a medium GPT-5 reply at 18.35 Wh on average.
A Comparison Table of Energy Usage
| Model / estimate | Energy per query | Who measured it | Real-world equivalent | Notes & caveats |
|---|---|---|---|---|
| “Old” ChatGPT estimate | ~3 Wh | 2023 third-party estimate | Incandescent bulb ~3 min | Now considered too high; based on dated hardware assumptions |
| GPT-4 | ~2.12 Wh | Univ. of Rhode Island, 2025 | Incandescent bulb ~2 min | Estimated; OpenAI has not published official per-model data |
| GPT-4o | ~0.3 Wh | Epoch AI, Feb 2025 | ~1 Google search | Ordinary text query; long inputs cost much more |
| Average query (OpenAI) | 0.34 Wh | Sam Altman blog, 2025 | Oven for ~1 second | Not peer reviewed; “average” undefined; excludes training |
| GPT-5 | ~18.35 Wh (up to 40) | Univ. of Rhode Island, Aug 2025 | Incandescent bulb ~18 min | Reasoning mode can raise draw 5 to 10x; hardware assumed |
Same chatbot, a roughly 60x spread depending on which model answers.
One detail matters more than the model name, which is the length of what you send and ask for. Epoch AI notes that a long input of around 10,000 tokens pushes a single query toward 2.5 Wh, and a 100,000-token input can approach 40 Wh on its own. Reasoning models that “think” before answering generate far more tokens, so a single complex prompt can cost many times an ordinary one. The headline numbers describe a short chat, not a 50-page document analysis.
ChatGPT Energy Use Per Day and Per Year
OpenAI says ChatGPT serves more than 2.5 billion queries a day for roughly 700 million weekly users. At Altman’s 0.34 Wh per query, that is about 850 megawatt-hours every day, and close to 310 GWh across a year. IEEE Spectrum puts that annual total at roughly the electricity used by 29,000 US homes, which sounds enormous until you set it against a single large data center or a mid-size city. The per-query cost is tiny, the aggregate is real, and both things are true at once.
The number changes sharply if heavier models take over the workload. If every one of those 2.5 billion daily queries ran on GPT-5 at the University of Rhode Island estimate, daily consumption could reach about 45 GWh. That is on the order of 1.5 million US homes, or the output of two to three nuclear plants. That is a worst-case scenario, not today’s reality, since OpenAI routes many simple requests to lighter models. It does show why model choice, not user count alone, drives the curve. As we covered in our look at AI trends and forecasts for 2026, electricity demand is becoming the industry’s real constraint.
Does ChatGPT Use More Energy Than a Google Search?
You have probably read that a ChatGPT query uses ten times the energy of a Google search. That comparison no longer holds. It traces back to a 2023 third-party estimate that pegged a ChatGPT query at about 3 Wh, set against Google’s own 2009 figure of roughly 0.3 Wh per search. The 3 Wh number assumed older hardware and longer outputs than ChatGPT actually produces.
When Epoch AI redid the math in 2025 with realistic token counts and modern GPUs, a typical GPT-4o query came out near 0.3 Wh, essentially the same as that decades-old Google search number. So for an ordinary text question, ChatGPT and a Google search are now in the same ballpark, not 10x apart. The caveat is that a heavy GPT-5 or reasoning query breaks the comparison again, since it can use tens of watt-hours. This is one of the most stubborn common myths about AI, and the underlying figure has quietly moved.
Why Newer ChatGPT Models Use More Energy
Two costs sit behind every answer, which are training and inference. Training is the one-time, very large cost of building the model; inference is the small repeated cost of answering your prompt. The per-query numbers above are inference only. A newer model is far hungrier per answer because it is bigger, generates more tokens, and increasingly “reasons” through extra hidden steps before replying.
GPT-4: ~2.12 Wh per query
University of Rhode Island researchers estimate GPT-4 at about 2.12 watt-hours per query, roughly an incandescent bulb running for two minutes. That is the baseline the newer numbers are measured against. It is an estimate, since OpenAI has not released official per-model energy data since GPT-3.
GPT-4o: ~0.3 Wh per query
The GPT-4o generation went the other way. Epoch AI’s 0.3 Wh estimate reflects more efficient modern GPUs plus a realistic, shorter token count than the original 2023 estimate assumed. For a stretch in 2025, ChatGPT looked leaner per query, which is the good-news part of the story.
GPT-5: ~18.35 Wh per query, up to 40
The University of Rhode Island estimate puts GPT-5 at about 18.35 Wh on average and up to 40 Wh for a medium 1,000-token answer, roughly 8.6x a GPT-4 query. The researchers had to assume the GPU setup since OpenAI does not disclose it. They also note GPT-5’s mixture-of-experts design lowers cost for some queries, while its reasoning mode can raise power draw five to ten times. The trend is clear even with the uncertainty, where smarter and more agentic models cost more per answer. The same dynamic shows up when AI agents chain many model calls together to complete a task.
Does Your Personal ChatGPT Use Actually Matter?
For an individual, the honest answer is that one ChatGPT query is a rounding error against a hot shower, a short drive, or streaming video for an evening. At 0.34 Wh, you could run hundreds of queries before matching the energy of boiling a single kettle. The footprint that matters is the aggregate one, driven by billions of users and by which model answers each request, not by whether you personally asked one extra question.
Your individual footprint is noise. The aggregate is the story.
The one lever you control is model choice. Firing the heaviest reasoning model at “rewrite this sentence” is the energy equivalent of taking a truck to buy milk. This is where a multi-model app like Fello AI helps. It gives you Claude, ChatGPT, Gemini, Grok and DeepSeek in one place for a single $9.99/month price, so you can match a lighter model to a simple task instead of defaulting to the biggest one. If you want to compare how the major assistants stack up on quality and efficiency, our ChatGPT vs Gemini comparison breaks down where each one earns its energy. The infrastructure question, covered in our piece on Nvidia’s $100B investment in OpenAI, is what will decide whether these numbers climb or level off.
The Bottom Line on ChatGPT’s Energy Use
A typical ChatGPT query uses about 0.34 watt-hours, close to a single Google search and far less than the scary headlines suggest, but the trend line points up as reasoning-heavy models like GPT-5 take over. The data-honest takeaway is that your individual use is negligible, the aggregate is meaningful, and model choice is the only real lever. Treat the single official number with healthy skepticism, since it is not peer reviewed and excludes training, and watch the per-model figures instead. If you use AI daily, the practical move is to pick the lightest model that does the job and let a multi-model tool handle the switching.
FAQ
How much energy does one ChatGPT query use?
About 0.34 watt-hours for an average query per OpenAI’s Sam Altman, or roughly 0.3 Wh for a typical GPT-4o query per Epoch AI. Heavy GPT-5 or reasoning queries can use 18 to 40 Wh.
How much energy does ChatGPT use per day?
At about 2.5 billion queries a day and 0.34 Wh each, ChatGPT draws roughly 850 MWh per day. If those queries ran on GPT-5, that could rise toward 45 GWh a day.
Does ChatGPT use more energy than a Google search?
Not really, not anymore. A modern GPT-4o query is about 0.3 Wh, roughly the same as Google’s own 0.3 Wh search figure. The old “10x” claim came from an outdated 2023 estimate.
Why does GPT-5 use so much more energy than GPT-4?
It is a larger model that generates more tokens and adds hidden reasoning steps. University of Rhode Island researchers estimate about 18.35 Wh per GPT-5 query versus 2.12 Wh for GPT-4, roughly 8.6x more.
Is the 0.34 watt-hours figure reliable?
Treat it as a useful but rough datapoint. It comes from OpenAI’s CEO, is not peer reviewed, does not define an “average” query, and excludes the energy used to train the model. Independent estimates broadly agree for ordinary queries but diverge sharply for heavy ones.
