How much CO₂ does one ChatGPT query produce?

Estimates from public research range from 0.3 to 5 grams of CO₂e per query, depending on model size, output length and the carbon intensity of the data centre's electricity supply. The calculator returns a model-specific figure.

Why does region matter?

The same compute uses very different amounts of fossil-based electricity depending on where the data centre is. Iceland (~10 g/kWh) runs on hydro and geothermal; coal-heavy grids approach 800 g/kWh.

How accurate is this?

It's a rough order-of-magnitude estimate. Public energy-per-token figures vary by 5× depending on hardware, batching efficiency and routing. Use the result for awareness, not for compliance reporting.

🌱 AI Carbon Footprint Calculator

Estimate the kWh and grams of CO₂ produced by an AI prompt — by model and grid region.

Model

Input tokens

Output tokens

Grid region (carbon intensity)

Prompts per day (optional)

ℹ️ Order-of-magnitude estimate. Energy-per-token figures vary by hardware generation and batching. Carbon intensity uses 2024 averages from Ember and Climatiq.

About the AI Carbon Footprint Calculator

Every prompt sent to a hosted LLM consumes a small amount of electricity to power inference servers. Multiplied across millions of users and billions of queries, it adds up. The AI Carbon Footprint Calculator turns model + token-count + grid-region into an estimated kWh and grams of CO₂ equivalent (gCO₂e).

The formula

CO₂ = tokens × Wh-per-token × grid intensity

Energy per token by model class

Small (≈ 7B params): ~0.2 mWh / token (Llama 3 8B, Mistral 7B)
Medium (≈ 70B): ~1 mWh / token (Llama 3 70B, Claude Haiku)
Large (≈ 400B+): ~3 mWh / token (GPT-4o, Claude Sonnet)
Frontier (≈ 1T+): ~6 mWh / token (GPT-4 / Claude Opus class)

Output tokens cost more than input tokens (typically 2–3× because each output requires a forward pass). Figures are rough — 2024–2025 published estimates vary by 3–5×.

Carbon intensity by region (gCO₂e / kWh)

Iceland: 11 (geothermal + hydro)
France: 56 (nuclear)
UK: 230
EU average: 250
Germany: 380
US average: 387
Texas (ERCOT): 410
India: 720
Coal-heavy grid: ~820

Frequently Asked Questions

Is this an exact number? No — treat it as an order-of-magnitude estimate. Real-world variation depends on hardware, batching efficiency, model routing and exact regional carbon intensity at time of inference.

What about model training? Training emissions are amortised across a model's lifetime usage. For widely-used models, training contributes roughly 5–20% of total lifecycle emissions.

How does this compare to a Google search? A standard Google search produces about 0.2 g CO₂e. A short LLM prompt to a frontier model is roughly 5–25× higher.

Pair this with our AI Model Cost Calculator and Prompt Cost Calculator for a full sustainability + cost picture.

⚠️

Important Note: Energy-per-token and carbon-intensity figures change rapidly. Use this tool for awareness; consult provider-published sustainability reports for compliance reporting.

Related Calculators

AI Model Cost Calculator — Compare API pricing for 100+ LLMs including GPT-4o, Claude, Gemini and Llama. Calculate exact token costs.
AI vs Human ROI Calculator — Compare AI tool costs against human labour. Monthly savings, annual ROI and break-even with live model pricing.
Carbon Footprint Calculator — Estimate your annual CO₂ emissions from transport, home energy, diet and lifestyle. Compare to global averages.