AI Energy Usage by Model: GPT vs Claude vs Gemini
On this page
The honest answer to "how much energy does an AI model use?" is: it depends, mostly on the size and type of model. Two different models can differ by a factor of five or more in energy per token. Below is a transparent, model-class view of AI energy usage, using the same factors Ecoia applies when measuring real requests. All figures are estimates that vary by model and region.
How AI energy is measured
For chat models, energy scales with tokens, the chunks of text a model reads and writes. We estimate a per-token energy figure (Wh per token) for each model class, then multiply by the data-center overhead, or PUE, around 1.56, to account for cooling and power distribution. To turn energy into carbon we apply grid intensity (a US average near 0.395 kg CO2 per kWh), and into water we apply roughly 3.4 liters per kWh across on-site cooling and off-site generation. The same method is detailed on our How it Works page.
Energy by model class
Rather than guess at individual product names, it is clearer to compare the four tiers that models fall into. Weight is shown relative to a mid-size model as the 1x baseline.
| Model class | Approx. Wh / token | Relative weight | Typical examples | Best for |
|---|---|---|---|---|
| Small / fast | ~0.0008 | ~0.5x | Haiku, mini, flash, nano variants | Drafting, summaries, classification, chat |
| Mid-size | ~0.0015 | 1x (baseline) | Balanced Sonnet / GPT-class models | Everyday reasoning, writing, coding help |
| Frontier / large | ~0.0038 | ~2.5x | Top-tier Opus / flagship models | Hard reasoning, long tasks, deep research |
| Reasoning-heavy | ~0.0040 | ~2.7x | o-series, R1, reasoning models | Multi-step math, planning, hard problems |
For image generation, energy is billed per image, not per token. We estimate around 2.0 Wh for a standard 1024x1024 image and about 3.5 Wh for a higher-quality image.
The headline: a frontier or reasoning model can use roughly five times the energy per token of a small one. Matching the model to the task is the easiest way to cut your footprint, as covered in how to reduce your AI carbon footprint.
Why larger and reasoning models cost more
Larger models have more parameters, so each token requires more computation to produce. Reasoning models compound this in two ways: they spend more energy per token, and they generate a large number of hidden reasoning tokens before they answer. A single hard question answered by a reasoning model can therefore cost several times the total energy of a quick reply from a small model, even though the visible answer looks similar in length.
Tokens vs images
Because chat is measured per token and images per image, comparisons need a common unit: energy. A standard image at roughly 2.0 Wh is in the ballpark of generating well over a thousand tokens from a mid-size model. Higher-resolution or higher-quality images cost more. So a handful of image generations can outweigh a long text conversation. If you generate many images, that is the part of your usage most worth measuring.
GPT vs Claude vs Gemini
People often ask which brand is greenest, but the brand is the wrong axis. Within every family there are small, mid-size and frontier models, and the tier dominates the energy picture. A small model from one provider will almost always beat a frontier model from another on energy per token. What actually varies between providers is the efficiency of their data centers and the carbon intensity of the grids they run on, which is rarely disclosed at the request level.
That is the gap Ecoia closes: instead of guessing, every request, whether GPT, Claude or Gemini, is measured for carbon, water and energy and then offset past 100%. You can estimate your own usage with the carbon footprint calculator, or read more on total demand in how much electricity does AI use.
FAQ
Which uses more energy, GPT, Claude or Gemini?
It depends far more on the model tier than the brand. A small model from any provider (such as a Haiku, mini, flash or nano variant) uses much less energy per token than a frontier or reasoning model. Comparing GPT vs Claude vs Gemini energy is really about comparing model classes, and exact figures vary by model and region.
Why do reasoning models use so much more energy?
Reasoning models think privately before answering, generating many extra internal tokens. At an estimated 0.0040 Wh per token, plus the larger volume of tokens they produce, the total energy for a single hard question can be several times that of a quick reply from a small model.
How much energy does generating an image use?
Images are billed per image rather than per token. We estimate roughly 2.0 Wh for a standard 1024x1024 image and around 3.5 Wh for a higher-quality image. A single image is therefore comparable to hundreds or low thousands of chat tokens, depending on the model.
Are these energy numbers exact?
No, they are estimates that vary by model architecture, hardware, batching and the region of the data center. We use them as transparent, comparable factors. Real-world figures from IEA analysis and provider sustainability reports vary, so treat all per-token numbers as approximate.