What is the most eco-friendly AI?

Among makers that publish figures, Google's median Gemini text prompt has the lowest reported footprint: 0.24 Wh and 0.03 g CO₂e (market-based). But the makers measure different things, so part of the gap is methodology. Claude, Grok, Meta AI and DeepSeek cannot be ranked at all, because their makers publish no per-prompt figures.

Is Claude more eco-friendly than ChatGPT?

Nobody outside Anthropic can say. Anthropic publishes no energy, emissions or water figures per prompt, so any comparison relies on independent estimates, which disagree widely.

Why is Mistral ranked last if it is the most transparent?

Because its figure measures more. Mistral published a full life-cycle analysis, including making the hardware, which naturally comes out higher than energy-only figures. Its rank reflects that difference in boundary, not necessarily a less efficient model — which is exactly why transparency is scored separately.

How can I use AI with less environmental impact?

Prefer text over generated images and video, which cost far more per item. Skip reasoning or deep-research modes for simple questions, since they generate many more tokens. Avoid regenerating answers you don't need. And keep perspective: diet, driving and flying move your footprint far more than prompts do.

The greenest AI models, ranked honestly

You can only rank what has been measured. So this page does two things: it ranks the models whose makers publish a per-prompt footprint, and it scores every maker on whether they publish anything at all.

Ranked by published CO₂e per text prompt

01
Gemini Google
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.
30 – 83 mg
0.24 Wh
02
ChatGPT OpenAI
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.
32 – 120 mg
0.34 Wh
03
Copilot Microsoft
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.
15 – 210 mg
0.16 – 0.6 Wh
04
Mistral Vibe (Mistral Large 2) Mistral AI
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.
1.1 g
energy not published

Ordered by the high end of each range. Ranges overlap, and Mistral's is a life-cycle figure — treat the order as indicative, not a verdict.

Cannot be ranked

Claude (Anthropic), Grok (xAI), Meta AI (Meta) and DeepSeek (DeepSeek) publish no per-prompt energy or emissions figures. Placing them anywhere in a ranking would be a guess dressed up as data.

Not disclosed

Transparency scorecard

Who publishes their AI's footprint?

A yes means the maker itself has published a per-prompt figure. It does not mean the figure is complete. Checked 2026-09-15.

Maker	Energy	Emissions	Water	Method	What exists
GeminiGoogle	Yes	Yes	Yes	Yes	A technical paper covering the full serving stack, reporting both market- and location-based emissions. Source
CopilotMicrosoft	Yes	No	Yes	Yes	Per-query energy for frontier-scale models in a peer-reviewed paper, including a reasoning scenario — not a figure for Copilot itself. Source
Mistral VibeMistral AI	No	Yes	Yes	Yes	A life-cycle analysis with Carbone 4 and ADEME: emissions, water and materials per response, but no energy figure. Source
ChatGPTOpenAI	Yes	No	Yes	No	A single figure in a blog post by its CEO, with no methodology, model or prompt length. Source
ClaudeAnthropic	No	No	No	No	No per-prompt figures found.
GrokxAI	No	No	No	No	No per-prompt figures found.
Meta AIMeta	No	No	No	No	No per-prompt figures found.
DeepSeekDeepSeek	No	No	No	No	No per-prompt figures found.

Using AI with less impact

01
Write, don't render
A generated image costs roughly as much as several text prompts, and a short video clip as much as thousands.
02
Save the heavy modes for heavy questions
Reasoning and deep-research modes generate far more tokens per answer. A simple question doesn't need them.
03
Don't regenerate for sport
Every regeneration is a new prompt. Edit the question instead of rolling the dice.
04
Reward the makers who publish
Transparency is the precondition for any of this to improve. It is a fair reason to choose one tool over another.
05
Keep perspective
Your food, travel and home energy are measured in tonnes. Start there.

Calculate your AI footprint Measure your whole footprint

Common questions

What is the most eco-friendly AI?: Among makers that publish figures, Google's median Gemini text prompt has the lowest reported footprint: 0.24 Wh and 0.03 g CO₂e (market-based). But the makers measure different things, so part of the gap is methodology. Claude, Grok, Meta AI and DeepSeek cannot be ranked at all, because their makers publish no per-prompt figures.
Is Claude more eco-friendly than ChatGPT?: Nobody outside Anthropic can say. Anthropic publishes no energy, emissions or water figures per prompt, so any comparison relies on independent estimates, which disagree widely.
Why is Mistral ranked last if it is the most transparent?: Because its figure measures more. Mistral published a full life-cycle analysis, including making the hardware, which naturally comes out higher than energy-only figures. Its rank reflects that difference in boundary, not necessarily a less efficient model — which is exactly why transparency is scored separately.
How can I use AI with less environmental impact?: Prefer text over generated images and video, which cost far more per item. Skip reasoning or deep-research modes for simple questions, since they generate many more tokens. Avoid regenerating answers you don't need. And keep perspective: diet, driving and flying move your footprint far more than prompts do.

Sources

Full per-model figures, including energy and water, are on our AI yardstick.