How to reduce your AI carbon footprint

By Kevin Naderian · Updated 17 November 2026

Short answer: The biggest wins: use efficient models for everyday tasks, save reasoning modes for hard problems, start fresh chats instead of very long ones, and go easy on images and video. A plain text prompt is only about 0.3 Wh.

First, where the energy goes

A language model does work for every token it reads and writes. A short text exchange is cheap (about 0.24–0.34 Wh by recent estimates). Costs rise with the size of the model, the length of the conversation it rereads each turn, hidden “thinking” tokens and media generation. So the tips below all aim at one thing: fewer, cheaper tokens for the same useful result.

9 practical tips

1. Right-size the model

You don't need the largest frontier model to rewrite an email. Efficient models, including mixture-of-experts designs that only activate part of the network per token, handle everyday tasks well for much less compute.

2. Save reasoning modes for hard problems

“Thinking” modes can generate thousands of hidden tokens before the answer. Use them for tricky maths, logic or code, not for a recipe.

3. Start a new chat for a new topic

Models reread the conversation on each reply. A 50-message thread about something else makes every new answer more expensive and often worse.

4. Write a clear first prompt

Give context, audience, format and length up front. One good prompt beats five regenerations.

5. Ask for the length you need

“In 3 bullet points” or “under 100 words” saves output tokens and your reading time.

6. Paste only what matters

Instead of a whole 40-page PDF, paste the relevant section.

7. Be intentional with images and video

Media generation uses far more energy than text. Generate when you need it, not for idle curiosity loops.

8. Batch related questions

“Give me a packing list, a 3-day itinerary and a budget” in one message avoids rereading context three times.

9. Choose transparent providers

Prefer companies that publish energy, water and methodology. It rewards transparency and helps you compare.

Which actions matter most?

ActionTypical impact
Avoiding unneeded video or image generationVery high
Skipping reasoning mode for simple tasksHigh
Using an efficient model for everyday chatHigh
Starting new chats instead of very long onesMedium
Clear prompts and concise answersLow to medium

These are relative rankings, not measurements. Estimate your own use with the AI carbon calculator.

A quick before-and-after

Before: one 80-message thread covering work, recipes and holiday plans, with reasoning mode always on and a few images generated “just to see”. After: separate short chats per topic, an efficient model by default, reasoning only for the tax question, and no throwaway images. Same useful answers, a fraction of the compute.

Keep perspective

For most people, AI text chat is a small slice of their footprint next to flights, driving, heating and diet. The point isn't guilt; it's good defaults. Read is AI bad for the environment? for the bigger picture.

How SproutGPT builds these in

  • Efficient open-weight models (DeepSeek, gpt-oss and Qwen) by default.
  • Deep Think reasoning is a Pro option you switch on, not the default.
  • Tight answers by default and short recent context.
  • Each reply shows exactly how many leaves it grew (about 1 per message).
  • Every 100 leaves earns a real tree from SproutGPT's revenue, tracked on the public ledger. Trees complement cutting emissions; they don't replace it.

Try SproutGPT free and watch exactly how many leaves each reply uses.

FAQ

What is the easiest way to reduce my AI carbon footprint?

Use an efficient model for everyday tasks, save reasoning modes for hard problems, and avoid unnecessary image or video generation.

Do long conversations use more energy?

Yes. Models reread the conversation context on each reply, so very long threads make every new answer more expensive. Start a new chat for a new topic.

Does being polite to AI waste energy?

Marginally. A few extra words add a few tokens. It's far less significant than model choice, reasoning modes or media generation.

Keep reading

Start growing your first tree 🌱

Free every day, no sign-up. Your first answer is a few seconds away.