"How much energy does ChatGPT use?" is one of the most-searched and most-contested questions about AI today — public estimates range from a fraction of a watt-hour to nearly 20 Wh per query depending on the model, the source, and what's included in the accounting. This calculator doesn't pretend the debate is settled. Instead, it puts your own personal AI use in perspective with transparent, sourced, editable assumptions, region-aware carbon accounting, and everyday equivalents you can actually picture.
Why the energy-per-query number is so contested
Estimates of AI energy use per query vary by an order of magnitude or more across public research, and for good reason: providers rarely publish exact figures, model architectures and serving infrastructure differ enormously, and a single 'prompt' can mean a two-word question or a multi-thousand-token reasoning chain. Early 2023-era estimates for GPT-3-class models cited numbers around 3 Wh per query; more recent analyses of efficient production models (like Google's disclosed Gemini figures and Epoch AI's ChatGPT estimates) suggest well under 1 Wh for typical short queries, while frontier reasoning models that "think" through many intermediate steps can use an order of magnitude more.
This calculator defaults to research-informed point estimates for five model classes — Small/efficient, Mid (GPT-class), Large frontier, Image gen, and Video gen — but exposes the Wh/prompt assumption as an editable input specifically because no single number is authoritative. If you have a better estimate for the model you actually use, plug it in.
Why your grid region changes the answer as much as your model does
A kWh of electricity does not carry the same carbon cost everywhere. The EPA's eGRID database shows US regional grid intensities ranging from roughly 0.08 kg CO2/kWh in the hydro-heavy Pacific Northwest to 0.45 kg CO2/kWh or more in coal-heavy regions — a nearly 6x spread. Nationally and internationally the spread is even wider: a nuclear- or hydro-dominant grid can emit well under 0.1 kg CO2/kWh, while a coal-dominant grid can exceed 0.9 kg CO2/kWh.
That means two people running the identical AI workload can have carbon footprints that differ by 5-10x purely because of where the electricity serving their queries was generated — which is why this calculator treats grid region as a first-class input, not an afterthought, and why the Model Comparison tab shows your region against a low-carbon grid alongside the model-class comparison.
Putting a small number in perspective — and what actually moves it
For the median personal use case — a few dozen chat prompts a day on a mainstream model — the annual footprint typically lands in the low single-digit kilograms of CO2, comparable to a handful of miles of driving. That's genuinely small next to categories like home energy use, air travel, or car commuting. But footprints scale linearly with usage, and the model-class multiplier is large: switching from an efficient model to a frontier reasoning model for the same task can multiply your footprint roughly 10x, and heavy image or video generation carries a substantially larger per-unit cost than text chat.
If you want to reduce your AI footprint, the two highest-leverage levers are the ones this calculator isolates: use the smallest model that gets the job done for routine tasks, and where you have a choice, prefer providers whose data centers run on cleaner grids. Everything else — turning off dark mode, closing browser tabs — is noise by comparison.