How Much CO2 Does ChatGPT Produce Per Query?
A single ChatGPT-style message produces a small but real amount of carbon dioxide, on the order of a fraction of a gram to a few grams of CO2, depending on the model, the length of the answer and where the data center runs. The number is tiny per message and meaningful at scale.
This guide shows how that figure is built up from first principles, using the same transparent factors Ecoia applies when it measures real requests. All numbers are estimates that vary by model and region.
How CO2 per query is calculated
Carbon from a chat request follows a short chain: tokens → energy → carbon. A model reads and writes tokens; each token costs a small amount of energy; that energy is produced by a grid with a known carbon intensity.
Ecoia estimates energy per token by model class, multiplies by data-center overhead (a PUE of about 1.56 for cooling and power distribution), then applies a grid intensity near 0.395 kg CO2 per kWh (a US average). The exact method is on the How it Works page.
A worked example
Take a mid-size model at roughly 0.0015 Wh per token. A 500-token answer (plus the prompt it read) might total around 700 tokens, or about 1 Wh at the chip, and roughly 1.6 Wh once you include the PUE overhead.
At 0.395 kg CO2 per kWh, 1.6 Wh works out to about 0.6 grams of CO2 for that message. A frontier or reasoning model can be several times higher; a small model, several times lower. See AI energy usage by model for the per-class figures.
Is that a lot?
Per message, no. A gram of CO2 is far less than sending a few emails or a minute of video streaming. The reason it matters is volume: hundreds of millions of queries per day multiply small numbers into data-center-scale demand, which is covered in how much electricity does AI use.
Image generation is the exception to watch. Images are billed per image, not per token, at roughly 2.0 Wh for a standard 1024x1024 image, so a handful of images can outweigh a long text chat.
How to lower it
The single biggest lever is matching the model to the task: use a small model for drafting and classification, and reserve frontier models for genuinely hard problems.
- Pick the smallest model that does the job well.
- Avoid regenerating answers repeatedly; refine the prompt instead.
- Batch related questions into one conversation rather than many cold starts.
- Prefer providers that measure and offset, so the residual footprint is neutralized.
The headline: A typical text answer is well under a gram to a few grams of CO2; model choice moves that by 5x or more, and images cost far more than text.
FAQ
How many grams of CO2 does one ChatGPT question produce?
A rough estimate is a fraction of a gram to a few grams for a text answer, depending on the model class, the answer length and the carbon intensity of the grid powering the data center. Frontier and reasoning models sit at the higher end; small models at the lower end.
Does a longer answer produce more CO2?
Yes. Energy scales with the number of tokens read and written, so longer prompts and longer answers use proportionally more energy and therefore more carbon.
How can I make my AI use carbon-negative?
Use a provider that measures each request and retires verified carbon and water offsets beyond 100% of the impact. Ecoia offsets past 200%, which makes each measured request net negative.