The short answer: a typical text prompt to ChatGPT or Gemini uses about 0.3 watt-hours of electricity, about as much as running a microwave for one second. The water depends on what you count: about 0.3 mL (five or six drops) evaporated to cool the data center, or about 1.3 mL once you add the water power plants use to make that electricity.
That's far less than the "bottle of water per email" figure that still circulates, and it isn't the whole story either. Long answers, pasted documents, "thinking" modes and image generation can multiply the footprint many times over. Here's what the best current measurements say, where the scary numbers came from, and what changes the footprint of your own prompts.
The numbers, from the people who measured them
| Source | What was measured | Energy per prompt | Water per prompt |
|---|---|---|---|
| Google (Aug 2025, data from May 2025) | Median Gemini text prompt, measured in production, including idle machines and cooling overhead | 0.24 Wh | 0.26 mL (on-site cooling only) |
| Sam Altman, OpenAI (Jun 2025) | "Average" ChatGPT query; no method published | 0.34 Wh | 0.32 mL (about 1/15 teaspoon) |
| Microsoft researchers (Joule, peer-reviewed, 2026) | Modeled median for frontier-scale models under production-like serving | 0.31 Wh (middle half: 0.16–0.60) | under 0.07 mL on-site (per press coverage) |
| Epoch AI (Feb 2025) | Modeled GPT-4o reply of about 400 words | about 0.3 Wh | not estimated |
| Mistral AI with ADEME and Carbone 4 (Jul 2025) | Full life cycle of a 400-token reply from a large model, including training and hardware | not disclosed | 45 mL |
The energy figures agree remarkably well: 0.24 to 0.34 Wh for a typical text prompt, from a company's own production data, a CEO's statement, a peer-reviewed study and an independent model. The water figures look wildly different, but that's because they measure different things.
Why the water numbers disagree: three different things called "water"
1. Cooling water at the data center
Data centers evaporate water to keep servers cool. This is what Google's 0.26 mL measures, and OpenAI's 0.32 mL appears to be the same kind of figure, though OpenAI didn't say. It's real, local and measurable, but it's the smallest slice.
2. Water used to make the electricity
Most power plants evaporate water too, especially coal, gas and nuclear plants that cool with water. Researchers at UC Riverside (Li et al., "Making AI Less Thirsty") estimate the US average at about 3.1 liters per kilowatt-hour of electricity. Add that to on-site cooling and a 0.3 Wh prompt comes to roughly 1.3 mL. Google's figure leaves this out, which critics like UC Riverside's Shaolei Ren have pointed out; Google's infrastructure sustainability lead called the researchers' grid assumptions flawed for Google's operations. Both sides have a point: it depends on where and when the prompt runs.
3. Life-cycle water, including training and hardware
Mistral's 45 mL is the broadest number published: it spreads the water used to train the model and manufacture the chips across every reply, for its large 123-billion-parameter model on France's grid. It's a valuable honest accounting of the whole system, but it isn't what a single extra prompt adds. Training water is already spent; sending one fewer prompt doesn't un-spend it.
So which number is right? For "what does my prompt add," the best answer is about 1.3 mL (cooling plus electricity), with the on-site 0.3 mL as the low end. That's the figure PromptSaver uses when it estimates the water your shorter prompts save.
Where "a bottle of water per email" came from, and why it's out of date
Three claims still get repeated in 2026, all built on early estimates:
"A bottle of water per 100-word email"
This came from a 2024 Washington Post analysis with UC Riverside (as reported by Tom's Hardware and others), which put a 100-word GPT-4 email at 519 mL of water and 0.14 kWh of electricity. That energy figure is about 400 times OpenAI's 0.34 Wh and about 580 times Google's measured 0.24 Wh. Analysts including Andy Masley and OpenAI engineer Shantanu Jain have traced how the estimate was inflated. Scale the water to today's measured energy and the email is closer to a milliliter than a bottle.
"A 500 mL bottle for every 20–50 questions"
This is UC Riverside's 2023 estimate for GPT-3, which assumed about 4 Wh per reply, 13 to 17 times today's measurements, and counted power-plant water. It was a reasonable estimate for 2023 hardware and software. It's often misquoted as "per question."
"ChatGPT uses 10 times the energy of a Google search"
This compared a 2023 estimate of about 3 Wh per AI query with Google's widely cited figure, from around 2009, of 0.3 Wh per search. Today's measured text prompt (0.24–0.34 Wh) is about the same as that old search number. Google says the energy for its median Gemini prompt fell 33 times in a single year, from May 2024 to May 2025, which shows how fast these numbers go stale.
The lesson isn't that AI's footprint doesn't matter. It's that per-prompt estimates from before 2025 are probably out of date, and even the 2025 measurements above will be superseded.
What actually changes the footprint of a prompt
The median prompt is small. Yours might not be. Here's what moves the number, biggest first.
Thinking or reasoning modes: about 7–30× more on average
Reasoning models write long hidden chains of thought before answering, and every one of those tokens uses energy. Hugging Face's AI Energy Score project found (as reported by SingularityHub) that reasoning modes used about 30 times more energy on average than the same models with reasoning off, with some cases far higher. Researchers measuring GPT-5 (Jegham et al.) estimated a medium prompt at about 2.3 Wh with minimal reasoning and about 17 Wh with high reasoning, close to a full phone charge for one answer.
Model size: about 5× or more
Bigger models use more energy per token. In the same study, a short prompt cost about 0.2 Wh on a small model and about 1.2 Wh on a large reasoning model. Mistral reports its impacts scale roughly in proportion to model size. Sending easy tasks to a smaller model is often the biggest single saving, for your bill and for the footprint.
Answer length: roughly proportional
Writing the answer is where most of the energy goes, because the model generates one token at a time. The ML.ENERGY benchmark found energy rises in proportion to response length. A 1,000-word answer uses roughly twice the energy of a 500-word one.
Pasting long documents: 8× to 100×+
Reading a long input costs energy too. Epoch AI estimates about 2.4 Wh for a 10,000-token input (about 7,500 words) and around 40 Wh at 100,000 tokens, compared with 0.3 Wh for a short question. Long chat histories count the same way, because every message re-sends the conversation.
Images and video: 2× to 3,000×
Generating an image takes roughly 0.6–3 Wh depending on the model and resolution (Luccioni et al.; MIT Technology Review), about as much as 2 to 10 text prompts. A few seconds of AI video can take hundreds of watt-hours, approaching 1 kWh for some models.
Location and time
The same prompt can use several times more water in one region than another, depending on the climate and on how local power is generated. UC Riverside's estimates ranged about 7× across US states and countries.
What it adds up to, in everyday terms
Using about 0.3 Wh and 1.3 mL per typical text prompt:
- A phone charge (19 Wh, per the US EPA) equals about 63 prompts.
- Boiling a liter of water in a kettle (about 0.15 kWh) equals about 500 prompts.
- A 10-watt LED bulb uses a prompt's worth of energy in about two minutes.
- One average US shower (about 60 liters) uses as much water as about 47,000 prompts, counting power-plant water, or about 200,000 counting cooling water alone.
- An average US home's daily electricity equals about 96,000 prompts.
Individually, a text prompt is tiny. At scale it isn't: billions of prompts a day add up, and the trend toward reasoning modes, agents and video is pushing the average up even as each model gets more efficient. Google, for example, still replenished only about 78% of its freshwater consumption in 2025.
How to shrink your own AI footprint (and your bill)
The habits that cut AI's footprint are the same ones that cut its cost:
- Ask for the length you need. Answer length drives energy almost one-for-one. "Five bullets, under 100 words" uses a fraction of an open-ended essay.
- Use a smaller model for simple tasks. A rewrite, a summary or a quick fact rarely needs a frontier model. PromptSaver names the smallest model that does each prompt well.
- Turn off thinking modes when you don't need them. They multiply energy per answer by about 7 to 30 times on average.
- Start a new chat for a new topic. Long conversations re-send everything, every message.
- Paste only the part of a document you need, and resize images before uploading.
- Get it right the first time. Every follow-up and redo is another full prompt. A clear prompt with the facts, format and length stated is the most efficient prompt there is.
PromptSaver shows the electricity and water a better prompt saves over a year, using the measured figures above (about 0.3 Wh per typical reply and 4.2 mL of water per Wh). It doesn't count training water, because a shorter prompt doesn't change it. For the money side of the same habits, see how to reduce your AI costs.
Frequently asked questions
How much water does one ChatGPT prompt use?
About 0.3 mL according to OpenAI, which appears to count data-center cooling only, or roughly 1.3 mL once you include the water used to generate the electricity. Life-cycle estimates that include training and hardware, like Mistral's 45 mL for a large model, are much higher.
How much energy does a ChatGPT query use?
About 0.3 watt-hours for a typical text query. OpenAI says 0.34 Wh on average, Google measured 0.24 Wh for its median Gemini prompt, and Microsoft researchers put the median at 0.31 Wh. Long answers, documents and reasoning modes use much more.
Does ChatGPT use more energy than a Google search?
For a typical text prompt today, about the same as Google's widely quoted 0.3 Wh per search, which dates from around 2009. The famous "10 times a search" figure came from a 2023 estimate; measured energy per AI prompt has fallen sharply since.
Is it true ChatGPT uses a bottle of water per email?
No. That figure came from a 2024 estimate that assumed about 400 times more energy per reply than OpenAI and Google now report. With today's measurements, a short email is closer to a milliliter than 500.
Why does AI use water at all?
Data centers evaporate water to cool their servers, and many power plants evaporate water to generate the electricity. Some data centers use air cooling or recycled water, so the amount varies a lot by location.
Do longer prompts use more energy?
Yes, but the answer matters more than the question. Energy rises roughly in proportion to the length of the reply. Pasting a long document also adds up: about 2.4 Wh for 10,000 tokens of input, per Epoch AI.
Do reasoning or thinking models use more energy?
Much more: about 7 to 30 times on average with reasoning switched on, across several studies, and some models far more. One estimate put a high-reasoning GPT-5 answer near 17 Wh, close to a full phone charge.
How much energy does it take to make an AI image?
Roughly 0.6 to 3 watt-hours per image, depending on the model and resolution, the same as about 2 to 10 text prompts. Short AI videos can take hundreds of watt-hours.
How we got these numbers
We used first-party disclosures from Google, OpenAI and Mistral, peer-reviewed and preprint research (Microsoft in Joule, UC Riverside, Epoch AI, ML.ENERGY, Jegham et al., Luccioni et al.) and the US EPA's equivalencies, checked in October 2026. Our 1.3 mL figure is our own calculation: 0.3 Wh times about 1.1 mL of cooling water per Wh (Google's measured ratio) plus about 3.1 mL of power-plant water per Wh (UC Riverside's US average). Everyday comparisons are our arithmetic from the cited sources. These figures are improving fast; we'll update this page as new measurements are published.