How we estimate AI prompt impact
LittleLeaf gives every prompt a green / amber / red rating from a transparent proxy model that runs entirely in the browser. Here is exactly how it works, and where its limits are.
1. Token estimate
We estimate the token count of a prompt from its word count: tokens = words ÷ 0.75. This is a deliberately simple, provider-agnostic approximation computed locally — the prompt text itself never leaves your browser.
2. Base energy per prompt
Each provider has a base energy figure (watt-hours per prompt), taken from published disclosures where they exist and a frontier-class proxy where they do not:
| Provider | Base (Wh) | Source |
|---|---|---|
| Google Gemini | 0.24 | Google Cloud (Aug 2025) |
| OpenAI ChatGPT | 0.34 | Altman / OpenAI (Jun 2025) |
| Anthropic Claude | 0.30 | Proxy — no disclosure |
| Microsoft Copilot | 0.34 | Proxy (as ChatGPT) |
| Perplexity | 0.28 | Proxy — mid-tier |
3. Complexity multiplier
Longer, more complex prompts cost disproportionately more energy. We scale the base figure by a multiplier keyed to the token estimate:
| Prompt size | Multiplier | Band |
|---|---|---|
| Under 200 tokens | 1.0× | Green baseline |
| 200–800 tokens | 2.5× | Amber range |
| Over 800 tokens | 5.0× | Red range |
This is grounded in Dauner et al. (2025), who found up to 6× more CO₂e for complex versus concise prompts, and Epoch AI (2025), who found energy rising roughly 13-fold at 15× token length.
4. The Environmental Impact Score
The score combines the three inputs with a time-of-use grid modifier:
EIS = base energy × complexity multiplier × grid modifier
The grid modifier reflects how clean the electricity grid is at the moment of the prompt, derived from hourly carbon-intensity data by region. The rating follows directly:
| Score | Rating | Meaning |
|---|---|---|
| EIS < 0.5 | Green | Sustainable prompt |
| EIS < 1.2 | Amber | Could be greener |
| EIS ≥ 1.2 | Red | High impact |
5. Water and carbon
Water use is estimated at roughly 1 mL per watt-hour, consistent with the disclosed Gemini (0.26 mL / 0.24 Wh) and ChatGPT (0.32 mL / 0.34 Wh) figures. Carbon is the energy figure multiplied by the regional grid carbon intensity (gCO₂e per kWh) at the time of use.
Uncertainty — read this
Sources
- Google Cloud (2025): 0.24 Wh, 0.26 mL per median Gemini text prompt.
- Altman / OpenAI (2025): 0.34 Wh, 0.32 mL per average ChatGPT query.
- Dauner et al. (2025), Frontiers in Communication: up to 6× more CO₂e for complex prompts.
- Epoch AI (2025): ~13× energy increase at 15× token length.