AI Carbon Footprint Calculator
Pick a model and how often you use it. See the energy, carbon and water it takes.
Your footprint
Light AI user · 6.1× a Light userPut it in perspective
Each scale is logarithmic: every step to the right is a much bigger jump. The shaded band is the likely range, and the white ring is a casual AI user. Drag a marker to try a different amount.
58% of a phone charge.
2% of a km by car.
10% of a glass of water.
Most per-query figures are estimates, so we show a likely range as well as a best guess. How we estimate →
Where your model sits
Every dot is one of 199 models, placed by the energy it uses per query on a log scale. Click a dot to try that model.
More detail — everyday comparisons, yearly total, breakdown, suggestions
In everyday terms
same electricity, per week- 23.3Google searches
- 5.3minutes of video streaming
- 8.4minutes of laptop use
- 0.06kettle boils
Each line is the same amount of electricity, not a sum. Reference values: web search 0.3 Wh, video streaming 0.08 kWh/hour, laptop 0.05 kWh/hour, kettle 0.11 kWh per boil.
The one change that would matter most
You're already on one of the lighter current options for this kind of tool, so the model isn't where the savings are — how often and how long you use it is.
199+ models · figures from published research · How we estimate
Frequently asked questions
How much energy does ChatGPT use?
A typical ChatGPT query running GPT-4o uses about 0.43 Wh of electricity — roughly 1.4 times a Google search (~0.3 Wh). OpenAI's Sam Altman has cited an average of 0.34 Wh, and independent estimates from Epoch AI (~0.3 Wh) and Jegham et al. (0.42 Wh) all put it at a fraction of a watt-hour. Longer prompts cost more: a heavy 1,000-token exchange is closer to 1.2 Wh, and very large contexts can reach tens of watt-hours. Lightweight models such as GPT-4o mini use roughly a tenth as much.
How much water does AI use per query?
Most of the water goes to cooling the data centres that run the model. A single short query uses only a few millilitres, but it adds up: independent research suggests a typical ChatGPT session of 10–50 messages can evaporate on the order of 500 mL of water. The exact figure depends heavily on the data centre's location and cooling design — arid regions and older evaporative cooling use far more per query than water-free or cold-climate sites.
Is AI bad for the environment?
Any single query is tiny — comparable to a Google search or a few seconds of a lightbulb. The concern is scale. AI and data centres consumed about 485 TWh of electricity in 2025 and are projected to reach 950 TWh by 2030, alongside rising water use for cooling. Spread across billions of daily queries, individual choices — using smaller models, keeping prompts short, and favouring cleaner-grid regions — compound into a meaningful difference.
Which AI model is the most energy-efficient?
Efficiency varies enormously: across the 199 models tracked here, the heaviest uses over 160,000 times more energy per query than the lightest. Smaller, distilled models — such as GPT-4o mini, Gemini Flash and Claude Haiku — typically use a small fraction of their flagship siblings' energy while handling most everyday tasks just as well. Browse the directory sorted by energy to compare models like for like.