A single short ChatGPT prompt uses roughly 0.3 watt-hours of electricity on a standard text model, going by the most-cited public estimates. That is about the same as running a modern LED bulb for a few minutes, or a small fraction of what your phone charger draws overnight. The carbon and water that come with it are small per prompt too. The honest caveat: these are estimates with wide error bars, not settled measurements, and the real figure for any given prompt could be several times higher or lower.
You can plug your own usage into the AI carbon footprint calculator to see energy, CO2, and water totals for a day, a month, or a whole team.
What one prompt actually costs
Three things get consumed when you send a prompt: electricity to run the chips, carbon from generating that electricity, and water to cool the data center. Rough per-prompt ranges for a short text reply look like this:
| Resource | Typical estimate (short reply) | What it compares to |
|---|---|---|
| Energy | ~0.3 Wh (range 0.1–3 Wh) | An LED bulb for a few minutes |
| Carbon | ~0.1–0.5 g CO2 | A few seconds of driving |
| Water | ~2–50 mL | A sip to a small glass |
The carbon number swings mostly on where the data center is. A prompt served from a grid running on hydro or nuclear power carries a fraction of the carbon of one served from a coal-heavy grid. Same model, same prompt, very different footprint.
Why the published numbers vary so much
If you search for this, you will find figures that disagree by 10x or more. That is not sloppiness, it reflects real uncertainty:
- Response length dominates. Energy scales with the number of tokens generated. A one-line answer and a 1,000-word essay differ by an order of magnitude.
- Model size matters. A small, fast model costs far less per token than a large frontier model. Providers rarely say which model variant served your request.
- Hardware and efficiency keep changing. Newer chips do more work per watt, so a 2024 estimate overstates a 2026 request.
- Providers do not publish exact figures. OpenAI, Google, and Anthropic disclose almost nothing per-request. Independent researchers reverse-engineer estimates from hardware specs and rough usage, so every number is a model of a model.
Treat any single “X watt-hours per prompt” claim as a midpoint of a wide range, not a fact.
Why reasoning models use far more
Reasoning models (the ones that “think” before answering) are the big exception to the small-footprint story. Before you see a single word, the model can generate thousands of hidden reasoning tokens. Since energy tracks tokens produced, a reasoning answer often uses 5 to 20 times the energy of a plain reply to the same question.
If most of your work runs through a reasoning model, your footprint is not 0.3 Wh per prompt, it is closer to a few watt-hours. That is the difference between a handful of Google searches and running your microwave for ten seconds. Our guide on why reasoning models cost more covers the token mechanics behind both the money and the energy.
Worked example: a month of moderate use
Say you send 40 prompts a workday, 20 workdays a month, so 800 prompts. Mix: 600 short standard replies and 200 reasoning replies.
- Standard: 600 × 0.3 Wh = 180 Wh
- Reasoning: 200 × 3 Wh (a mid estimate) = 600 Wh
- Monthly total: ~780 Wh, roughly 0.78 kWh
That 0.78 kWh is about what a typical fridge uses in a day, or the cost of a couple of loads in an efficient dryer. For one person’s month of AI, small. Multiply by a 50-person team and you are at ~39 kWh a month, which is worth measuring but still modest against the team’s laptops, screens, and video calls. Run these numbers for your own volume in the AI carbon footprint calculator.
How to keep it low
- Prefer smaller models for routine tasks and save reasoning models for problems that need them.
- Ask for shorter answers when you do not need an essay. Fewer output tokens, less energy.
- Batch related questions into one conversation instead of re-sending context repeatedly.
- If you run your own models, hosting location and hardware efficiency matter more than almost anything else, which is where the self-hosting vs API calculator helps you weigh the tradeoff.
The takeaway: one prompt is genuinely small, reasoning and long outputs are where it adds up, and every number you see is an estimate. Use the ranges to reason about scale, not to quote a precise footprint.
