Is ChatGPT bad for the environment?

Per question, not much. OpenAI says an average ChatGPT query uses about 0.34 watt-hours — roughly what an LED bulb uses in a couple of minutes — though it has published no methodology behind that figure. The environmental concern is scale: the same small number multiplied across a very large user base, plus image and video generation, which cost far more per item.

How much energy does one AI prompt use?

The makers that publish figures put a typical text prompt between about 0.16 and 0.6 watt-hours: Google measured 0.24 Wh for a median Gemini prompt, OpenAI states 0.34 Wh for ChatGPT, and Microsoft measured a median of 0.31 Wh for a frontier-scale query. Reasoning modes use far more — Microsoft puts a long reasoning answer at 4.32 Wh.

How much water does AI use?

Google reports 0.26 mL of water for a median Gemini prompt and OpenAI about 0.32 mL for ChatGPT. Mistral's life-cycle study, which counts much more than the data centre, arrives at 45 mL for a 400-token response. Anthropic, Meta, xAI and DeepSeek publish no water figures.

Which AI model is the most eco-friendly?

Nobody can honestly rank them yet, because most makers publish nothing and the ones that do measure different things. Among published figures, Google's median Gemini prompt is the lowest, but it is measured differently from Mistral's life-cycle figure, so the gap is partly methodology rather than efficiency.

Is generating AI images or video worse than text?

Yes, by a lot. Independent measurements put a single generated image at roughly 0.6 to 3 watt-hours and a few seconds of video at roughly 30 to 940 watt-hours — up to thousands of text prompts for one clip.

Can I make my AI use carbon neutral by offsetting it?

No — and anyone who promises that is overclaiming. Funding verified carbon credits supports climate projects equal to that amount, but it does not undo the emissions themselves. What you can honestly say is how many kilograms of credits you funded.

Is AI bad for the environment?

Per prompt, far less than most people fear. Per image and per video clip, a lot more. And for most AI makers, nobody outside the company actually knows — because they publish nothing.

This page is our running yardstick: every figure labelled with who produced it, shown as a range, with the source one click away.

A text promptpublished range, large models: 0.16 – 0.6 Wh
An AI imageindependent measurement: 0.63 – 2.9 Wh
A short AI videoindependent measurement: 30 – 940 Wh

Calculate yours

What does your AI use add up to?

Which AI do you use most?

Text prompts

per day

Images generated

per week

Video clips generated

per month

Your AI use, per year

230 – 860 gCO₂e

Less than a tenth of a quarter-pound beef burger.

Text230 – 860 g
Images—
Video—

Water from text prompts: 2.3 L a year. Nobody publishes water figures for images or video.

ChatGPT: disclosed by the maker. An average ChatGPT query, as stated by OpenAI's CEO. OpenAI has published no methodology, model or prompt length behind it, and no emissions figure — the CO₂e range is our calculation from the energy figure.

For most people that is a rounding error. A typical footprint is measured in tonnes, and diet, driving, flying and home energy are where the tonnes come from.

Measure your whole footprint Find the pass for your AI plan

The yardstick

One text prompt, model by model

The square tells you who the number comes from. Where a maker publishes energy but not emissions, we convert it using the grid range Google itself reports (94–345 g CO₂e per kWh).

Model	Source	Energy	CO₂e	Water
GeminiGoogle	Disclosed by the maker	0.24 Wh	30 – 83 mg	0.26 mL
ChatGPTOpenAI	Disclosed by the maker	0.34 Wh	32 – 120 mg	0.32 mL
Mistral Vibe (Mistral Large 2)Mistral AI	Disclosed by the maker	—	1.1 g	45 mL
CopilotMicrosoft	Disclosed by the maker	0.16 – 0.6 Wh	15 – 210 mg	0 – 0.067 mL
ClaudeAnthropic	Not disclosed	0.16 – 0.6 Wh	15 – 210 mg	—
Grok, Meta AI, DeepSeek & othersVarious	Not disclosed	0.16 – 0.6 Wh	15 – 210 mg	—

Gemini: The median Gemini Apps text prompt, measured across the full serving stack — AI chips, host machines, idle capacity and data-centre overhead. The low CO₂e figure is Google's own (market-based); the high end applies Google's location-based grid factor, which is our calculation. Source
ChatGPT: An average ChatGPT query, as stated by OpenAI's CEO. OpenAI has published no methodology, model or prompt length behind it, and no emissions figure — the CO₂e range is our calculation from the energy figure. Source
Mistral Vibe (Mistral Large 2): One 400-token response from Mistral Large 2 in Le Chat (since renamed Mistral Vibe), from a life-cycle analysis carried out with Carbone 4 and the French ecological transition agency ADEME. A life-cycle figure counts far more than electricity — including making the hardware — so it runs higher and is not directly comparable to the energy-only figures above. Source
Copilot: Microsoft's published figure for a typical query of about 300 output tokens to a frontier-scale model: median 0.31 Wh, with the middle half of queries between 0.16 and 0.60 Wh. Not a figure for Copilot specifically; CO₂e is our calculation. Source Source
Claude: Anthropic publishes no per-prompt energy, emissions or water figures. We show Microsoft's measured range for a frontier-scale query as a stand-in. Reasoning-heavy use costs far more: the same study puts a ~5,000-token reasoning answer at 4.32 Wh. Source
Grok, Meta AI, DeepSeek & others: None of these makers publish per-prompt figures. We show Microsoft's measured range for a frontier-scale query as a stand-in. Reasoning modes generate many more tokens per answer, so real use can sit well above this range. Source

See the ranking and which makers publish nothing →

For scale

One beef burger, in prompts

A quarter-pound patty from a beef herd comes to about 9.6 kg CO₂e — from the same food data we publish. Here is how many text prompts it takes to match it.

This is not an argument that AI is harmless. It is a reason to spend your attention where the tonnes are.

Gemini Disclosed by the maker120,000 – 320,000
ChatGPT Disclosed by the maker82,000 – 300,000
Mistral Vibe (Mistral Large 2) Disclosed by the maker8,400
Copilot Disclosed by the maker46,000 – 640,000
Claude Not disclosed46,000 – 640,000
Grok, Meta AI, DeepSeek & others Not disclosed46,000 – 640,000

Where it gets heavy

Images and video cost far more than text

One imageIndependent estimate

0.63 – 2.9 Wh

60 mg – 1 g CO₂e

Measured on open image models, since the commercial generators publish nothing. The spread covers efficient modern models up to the average across models tested.

One short clipIndependent estimate

30 – 940 Wh

2.8 – 330 g CO₂e

A few seconds of video from an open video model, measured by MIT Technology Review. Commercial video models publish nothing; longer or higher-resolution clips cost more.

Why the numbers disagree

Different boundaries.Google, OpenAI and Microsoft report the electricity a query uses in the data centre. Mistral's figure is a life-cycle analysis that also counts manufacturing the hardware. Neither is wrong; they answer different questions.

Different grids.The same watt-hour emits less if you count a company's clean-energy contracts (market-based) than if you count the grid its data centres actually draw from (location-based). Google publishes both; we show both.

Different prompts.A "median prompt" is short. A long document, a coding session or a reasoning model thinking step by step can use many times more.

Silence. Anthropic, Meta, xAI and DeepSeek publish no per-prompt figures. Where we show a number for them it is a labelled stand-in, not a measurement.

Common questions

Is ChatGPT bad for the environment?: Per question, not much. OpenAI says an average ChatGPT query uses about 0.34 watt-hours — roughly what an LED bulb uses in a couple of minutes — though it has published no methodology behind that figure. The environmental concern is scale: the same small number multiplied across a very large user base, plus image and video generation, which cost far more per item.
How much energy does one AI prompt use?: The makers that publish figures put a typical text prompt between about 0.16 and 0.6 watt-hours: Google measured 0.24 Wh for a median Gemini prompt, OpenAI states 0.34 Wh for ChatGPT, and Microsoft measured a median of 0.31 Wh for a frontier-scale query. Reasoning modes use far more — Microsoft puts a long reasoning answer at 4.32 Wh.
How much water does AI use?: Google reports 0.26 mL of water for a median Gemini prompt and OpenAI about 0.32 mL for ChatGPT. Mistral's life-cycle study, which counts much more than the data centre, arrives at 45 mL for a 400-token response. Anthropic, Meta, xAI and DeepSeek publish no water figures.
Which AI model is the most eco-friendly?: Nobody can honestly rank them yet, because most makers publish nothing and the ones that do measure different things. Among published figures, Google's median Gemini prompt is the lowest, but it is measured differently from Mistral's life-cycle figure, so the gap is partly methodology rather than efficiency.
Is generating AI images or video worse than text?: Yes, by a lot. Independent measurements put a single generated image at roughly 0.6 to 3 watt-hours and a few seconds of video at roughly 30 to 940 watt-hours — up to thousands of text prompts for one clip.
Can I make my AI use carbon neutral by offsetting it?: No — and anyone who promises that is overclaiming. Funding verified carbon credits supports climate projects equal to that amount, but it does not undo the emissions themselves. What you can honestly say is how many kilograms of credits you funded.

Sources

Last reviewed 2026-09-15. When a maker publishes a new figure, this page changes — tell us at thijnfelix@carbonshredder.com.

Now the number that matters

Your AI use is one line. Your whole footprint takes two minutes to measure.

Calculate my footprint Explore the food data