The footprint of a prompt
Large language models cost energy, water, and carbon to run. But how much? Pick an everyday activity below and see how many AI prompts — or tokens — it would take to match its carbon footprint.
That's about — of CO2.
To match that with AI, you'd send roughly
—
text prompts
≈ — tokens (at ~500 tokens each)
Assumptions & sources (edit the numbers)
Per AI prompt
Casual uses Google's median Gemini Apps text prompt ("comprehensive" methodology, Aug 2025): 0.24 Wh, 0.26 mL water, 0.03 gCO2e, ~500 output tokens. These are first-party estimates using favorable boundaries (median not mean, text-only, market-based carbon, training excluded) — a best-case lower bound. Independent figures agree on the order of magnitude: Epoch AI ~0.3 Wh/query, Sam Altman 0.34 Wh/query.
Professional models a heavy reasoning / long-context request at roughly 40× the casual prompt (~9.6 Wh, ~10 mL, ~1.2 gCO2e, ~8,000 tokens). Epoch AI estimates a 10k-token input at ~2.5 Wh and a 100k-token query at ~40 Wh; reasoning models also emit thousands of hidden reasoning tokens per answer, so this is a deliberately heavier — but still conservative — profile. Picking a profile just presets the numbers below; edit any of them to taste.
Sources
- Google — Measuring the Environmental Impact of AI Inference (Aug 2025): 0.24 Wh / 0.26 mL / 0.03 gCO₂e per median prompt.
- Epoch AI — How much energy does ChatGPT use? (Feb 2025): ~0.3 Wh and ~500 output tokens per query.
- US EPA — GHG from a typical passenger vehicle: ~400 g CO₂ per mile.
- US EPA — Equivalencies Calculator references: 5.3 kg CO₂ per therm of natural gas.
- Flight & per-day heating figures are reasonable estimates, not independently verified here — adjust them above to taste.
Want the full picture? Read the deep research report behind these numbers — every claim, source, and caveat.