How Much Energy Does ChatGPT Use per Query? | The Real Math

Most people asking this want one number. The honest answer is a range.

For a plain text prompt and a normal-length reply, the best current public estimates put one ChatGPT query at roughly 0.3 to 0.34 watt-hours. That is far lower than the older 2.9 to 3 watt-hour figure that still gets repeated in blog posts and social posts.

That gap is not a small detail. It changes the whole story. If you quote the old number, you make ChatGPT look about ten times more power-hungry than newer estimates suggest. If you quote the newer number as if it covers every task, you miss how fast usage climbs when prompts get long, files are uploaded, or the model spends extra compute on reasoning.

So the clean answer is this: most everyday text queries are probably in the few-tenths of a watt-hour range, not multiple watt-hours. But there is no single fixed cost for every prompt.

What Changes ChatGPT Energy Per Query In Real Use

The number moves because a “query” is not one standard unit. One person asks for a two-line rewrite. Another asks for code, uploads a PDF, and requests a long answer with several passes of reasoning. Those are not the same job.

Prompt Length And Reply Length

More tokens mean more work. A short prompt with a short reply needs less compute than a long back-and-forth thread or a huge pasted document. That is why a simple one-line question can be cheap while a dense research request costs more.

Model Choice

Different models have different serving costs. Newer systems can be more efficient per token, but some higher-effort modes also spend more compute to produce stronger answers. That means a newer model can be cheaper in one case and costlier in another.

Reasoning And Tool Use

Some replies are close to plain next-word prediction. Others call tools, search, or spend extra time on internal reasoning. Those extra steps can push energy use above the average text-only estimate.

Data Center Setup

The same prompt can carry a different energy and water cost depending on where it runs. Hardware generation, server load, cooling method, and the local power mix all matter. That is one reason older one-number claims aged badly.

OpenAI’s public figure came from Sam Altman’s June 2025 post, where he said an average query uses about 0.34 watt-hours. A separate Epoch AI estimate for GPT-4o landed at roughly 0.3 watt-hours for a typical query. Those two numbers are close enough to treat them as the current center of gravity for normal text use.

Why The Old 3 Wh Number Still Shows Up

The older estimate came from early model-era assumptions, older hardware, and rough public math. It was useful when companies shared little data, but it stuck around long after the market changed.

That old figure still matters in one way: it reminds you that averages drift fast. AI systems are getting cheaper to serve, but that does not mean every prompt is cheap. A terse chat reply and a long reasoning job can sit very far apart.

So if you see one sentence claiming that ChatGPT “uses 3 Wh per question,” treat it as old context, not a solid current default.

How Much Energy Does ChatGPT Use per Query? The Best Current Range

If you want a practical number to use in plain English, use this:

• Typical text prompt: about 0.3 to 0.34 Wh
• Lean, short prompt: often lower than that
• Long, tool-heavy, or reasoning-heavy task: can climb well above that

That range also lines up with another fresh data point outside OpenAI. Google said the median Gemini Apps text prompt uses 0.24 Wh. Different systems are not identical, but that figure still helps show where newer large-model inference can land when the workload is plain text rather than a heavier task.

Estimate Or Case	Energy Use	What It Means
OpenAI average query claim	0.34 Wh	Public 2025 average for a ChatGPT query
Epoch AI typical GPT-4o query	~0.3 Wh	Fresh estimate for ordinary text use
Google Gemini median text prompt	0.24 Wh	Another fresh benchmark for text-only inference
Older widely shared estimate	2.9 to 3 Wh	Early-model-era math still quoted online
Short prompt, short reply	Below average	Less token processing and less compute time
Long prompt or long reply	Above average	More tokens means more server work
Reasoning-heavy task	Can rise sharply	Extra compute pushes the query cost up
File-heavy or tool-using task	Can rise sharply	Extra steps go beyond plain text generation

What 0.3 Wh Actually Looks Like

A watt-hour is a tiny unit, so the number can feel abstract. Here is the plain reading: one normal ChatGPT prompt is not a household-scale event. It is small on its own. The bigger issue shows up when billions of prompts stack together across huge data center fleets.

That is why both of these statements can be true at the same time:

• One ordinary query uses a small amount of electricity.
• AI can still drive major grid growth at scale.

The International Energy Agency’s Energy and AI work says global data center electricity use is on track to rise sharply through 2030. That growth is driven by volume, not by one person asking for help with a short email.

Energy Is Not The Whole Footprint

Energy gets the headline, but cooling and water use matter too. That part is messy because location matters a lot.

In June 2025, Sam Altman also said an average ChatGPT query uses about 0.000085 gallons of water, which is roughly one-fifteenth of a teaspoon. That is much lower than the older bottle-per-email headlines tied to GPT-4-era estimates for a 100-word email in certain data center conditions.

Those older stories were not fake. They were tied to a different setup, a different workload, and location-sensitive water math. What they do show is that “per query” numbers can swing hard depending on the model, the task, and where the servers sit.

When A Query Costs More Than The Average

You should assume the average breaks down in these cases:

• You paste in a long report, legal text, or codebase
• You ask for a long answer with many revisions
• You use a reasoning mode that spends more compute
• You generate images, audio, or other media
• You trigger web search, file reading, or other tools

That is why “per query” can be a slippery phrase. A one-line grammar fix and a multi-step research task are both queries, yet their server cost is nowhere near the same.

Query Type	Likely Relative Cost	Why
Short text question	Low	Few tokens and little processing time
Normal chat reply	Average	Fits the 0.3 to 0.34 Wh ballpark
Long drafting task	Medium to high	Longer output drives extra compute
Reasoning-heavy request	High	More internal compute before the reply
Multimodal or tool-using task	High	Extra systems run beyond plain text generation

So, What Number Should You Use?

If you are writing, teaching, or quoting a current estimate, use about 0.3 to 0.34 Wh per normal text query and say clearly that the figure is an average, not a fixed law.

If you want the safer wording, say that ChatGPT energy use per query depends on workload, with ordinary text prompts in the few-tenths-of-a-watt-hour range and heavier jobs costing more. That wording is accurate, current, and much harder to knock down.

One last point: this topic changes fast. Model serving gets cheaper, data centers change, and products add new tools. So the best number today is a fresh estimate, not a forever number.

For the latest public anchor, use Sam Altman’s June 2025 note on average query energy as OpenAI’s own public claim, then pair it with independent estimates so you do not lean on one company’s word alone.