Methodology & Sources
Transparency is key. Here is how we estimate the numbers.
Core Assumptions
Estimating the exact footprint of proprietary closed-source models (like GPT-4) is difficult because companies do not release exact energy data per query. However, using research papers and industry benchmarks, we can create high-confidence estimates.
Energy per Request
Energy consumption is based on model parameter estimates and PUE (Power Usage Effectiveness) overhead of 1.15x for data center cooling.
- GPT-4:~0.003 - 0.01 kWh / request
- GPT-3.5:~0.0003 kWh / request
- Images:~2.9 Wh / image avg (up to ~12 Wh for HD)
Note: Generating output tokens is ~3x more energy-intensive than processing input tokens due to autoregressive decoding cycles.
Carbon Footprint & Lifecycle
We convert Energy (kWh) to Carbon (gCO2e) based on grid intensity:
- Global Avg:~475 gCO2e/kWh
- Green Grid:~20 gCO2e/kWh (Hydro/Wind)
Optional: Hardware manufacturing (embodied carbon of servers and GPUs) adds an estimated +30% lifecycle overhead to operational emissions.
Water Consumption
Data centers consume water for cooling (Scope 1) and electricity generation (Scope 2). Research indicates a significant "water footprint" for AI.
"Making AI Less Thirsty" (Li et al., 2023) estimates that a conversation of 20-50 questions consumes ~500ml of water.
Reviewed 2026-10-07
AI model energy estimates
This is the canonical model registry used by the calculator and API. Values estimate inference energy before the calculator applies data-center overhead, grid carbon intensity and water usage. Provider catalogs confirm model availability; they do not supply these electricity figures.
| Model | Provider | Estimated energy | Output multiplier |
|---|---|---|---|
| GPT-6 Astra (Frontier) | OpenAI | 7.0 Wh / 1K tokens | 3.5× |
| GPT-5.6 Sol | OpenAI | 4.5 Wh / 1K tokens | 3.2× |
| GPT-5.6 Terra | OpenAI | 2.4 Wh / 1K tokens | 3.0× |
| GPT-5.6 Luna | OpenAI | 0.30 Wh / 1K tokens | 2.5× |
| OpenAI o1 (Reasoning) | OpenAI | 15.0 Wh / 1K tokens | 3.5× |
| OpenAI o1-mini (Reasoning) | OpenAI | 5.0 Wh / 1K tokens | 3.0× |
| OpenAI o3-mini (Reasoning) | OpenAI | 4.5 Wh / 1K tokens | 3.0× |
| GPT-5.5 Pro | OpenAI | 6.0 Wh / 1K tokens | 3.0× |
| GPT-5.5 | OpenAI | 3.0 Wh / 1K tokens | 3.0× |
| GPT-5.4 Pro | OpenAI | 5.0 Wh / 1K tokens | 3.0× |
| GPT-5.4 | OpenAI | 2.8 Wh / 1K tokens | 3.0× |
| GPT-4o | OpenAI | 3.5 Wh / 1K tokens | 3.0× |
| GPT-4o mini | OpenAI | 0.40 Wh / 1K tokens | 2.5× |
| Claude Fable 5.1 (Frontier) | Anthropic | 7.0 Wh / 1K tokens | 3.5× |
| Claude Opus 5.5 | Anthropic | 4.8 Wh / 1K tokens | 3.2× |
| Claude Sonnet 5.5 | Anthropic | 2.0 Wh / 1K tokens | 3.0× |
| Claude Opus 5 | Anthropic | 4.5 Wh / 1K tokens | 3.2× |
| Claude Sonnet 5 | Anthropic | 1.8 Wh / 1K tokens | 3.0× |
| Claude Sonnet 4.6 | Anthropic | 1.8 Wh / 1K tokens | 3.0× |
| Claude Opus 4.5 | Anthropic | 4.0 Wh / 1K tokens | 3.0× |
| Claude Haiku 4.5 | Anthropic | 0.40 Wh / 1K tokens | 2.5× |
| Claude 3.5 Sonnet | Anthropic | 2.0 Wh / 1K tokens | 3.0× |
| Gemini 3.1 Pro (Preview) | 4.0 Wh / 1K tokens | 3.2× | |
| Gemini 3.8 Flash | 0.45 Wh / 1K tokens | 2.8× | |
| Gemini 3.5 Flash-Lite | 0.12 Wh / 1K tokens | 2.2× | |
| Gemini 3.5 Flash | 0.30 Wh / 1K tokens | 2.5× | |
| Gemini 2.0 Flash | 0.35 Wh / 1K tokens | 2.5× | |
| DeepSeek V4 Pro | DeepSeek | 3.8 Wh / 1K tokens | 3.2× |
| DeepSeek V4 Flash | DeepSeek | 0.45 Wh / 1K tokens | 2.8× |
| DeepSeek-V3 (MoE) | DeepSeek | 1.2 Wh / 1K tokens | 3.0× |
| DeepSeek-R1 (Reasoning) | DeepSeek | 13.0 Wh / 1K tokens | 3.5× |
| Llama 4 Maverick | Meta (Open Source) | 0.35 Wh / 1K tokens | 2.5× |
| Llama 4 Scout | Meta (Open Source) | 0.30 Wh / 1K tokens | 2.5× |
| Llama 3 70B | Meta (Open Source) | 0.50 Wh / 1K tokens | 2.5× |
| Small Models (7B) | Meta / Community | 0.10 Wh / 1K tokens | 2.0× |
| Mistral Large 4 (Preview) | Mistral AI | 3.2 Wh / 1K tokens | 3.2× |
| Mistral Medium 3.5 | Mistral AI | 1.5 Wh / 1K tokens | 3.0× |
| Mistral Small 4 | Mistral AI | 0.35 Wh / 1K tokens | 2.5× |
| Mistral Large 3 | Mistral AI | 2.8 Wh / 1K tokens | 3.0× |
| Mistral Large | Mistral AI | 2.5 Wh / 1K tokens | 3.0× |
| Image Generation (GPT Image 2 / Nano Banana 2) | Various | 2.9 Wh / image | Not applicable |
Regional Water Impact (API v2)
Our v2 API uses region-specific data for each data center location. Water impact depends on three factors:
Water Usage Effectiveness (WUE)
Measures liters of water consumed per kWh of IT energy. Ranges from 0.3 L/kWh (Sweden) to 3.0 L/kWh (UAE).
Water Stress Classification
From WRI Aqueduct Water Risk Atlas. Classifies regions as low, moderate, high, or critical.
Cooling Technology
Evaporative cooling uses more water but less energy. Air cooling reduces water use in cooler climates.
Data Sources & References
WRI Aqueduct Water Risk Atlas
Global water stress mapping used for regional water impact classification.
View Atlas →EIA State Electricity Profiles
US Energy Information Administration data on state-level grid carbon intensity.
View Profiles →Corporate Sustainability Reports
Microsoft (2024) & Google (2025) sustainability reports.
Microsoft → Google →Hugging Face & CMU
Power Hungry Processing: Watts Driving the Cost of AI Deployment? (Luccioni et al., FAccT 2024) — our primary anchor for per-inference energy, including the ~2.9 Wh average per generated image.
Read Paper →From Prompts to Power
Caravaca, Cuevas & Cuevas (2025) — 32,500+ hardware measurements confirming that output tokens dominate energy and that batch size strongly affects per-prompt cost.
Read Paper →Google — AI at Scale
Measuring the Environmental Impact of Delivering AI at Google Scale (2025) — first per-prompt disclosure from a major lab: a median Gemini text prompt uses ~0.24 Wh.
Read Paper →Known Limitations
- Batch size. Our per-request figures model a single, unbatched inference. Production AI services process many prompts in parallel, which can lower the real per-prompt energy substantially (Caravaca et al., 2025, measured up to a ~36x reduction on large models). Treat our numbers as a transparent upper bound rather than the exact cost of a hosted API call.
- Closed-model estimates. Providers of proprietary models (GPT, Claude, Gemini) do not publish per-query energy. Those rates are extrapolated from measured open models by model tier, not directly measured.
- Hardware & grid variance. The same model can vary 2x or more by GPU, and grid carbon intensity changes by hour and location. We use representative averages, not real-time values.
Disclaimer: These figures are estimates intended for educational purposes. The AI landscape changes daily. We update our constants as new peer-reviewed data becomes available.