What is the carbon footprint of one ChatGPT prompt?

Published estimates for one AI prompt run from 0.02 g CO2e to a few grams for a short prompt, and up to about 17 g for a long prompt on a high-carbon host, and the spread comes mostly from the grid behind the data centre and from what each study counts. Google reports 0.03 g for its median Gemini Apps text prompt using a market-based factor of 94 g CO2e per kWh. Mistral AI reports 1.14 g for a 400-token Le Chat response in a lifecycle study that includes upstream impacts. These are estimates built on assumptions, not meter readings.

The table lists figures opened at their primary sources, with the boundary each publisher gives. None of them is a figure for ChatGPT itself, so each describes a different service, model or workload.

The 2023 row is the oldest and the only one that adds training. Its authors took the operating figure from a blog post, so it is better read as an order of magnitude than as a result. The rows from 2025 onward are narrower and newer.

Because the carbon is the electricity used times the carbon intensity of the grid supplying it. Berget AI makes the point with one model in two places. In Sweden the request comes to 0.38 g, of which 89% is hardware manufacturing. In Texas it comes to 3.2 g, of which 90% is electricity. The energy is about 4 Wh in both. A low-carbon grid shifts the footprint from the electricity to the hardware, and a high-carbon grid does the opposite.

SAEF shows the same pattern. For a class B model, one exchange in EU North - Sweden comes to 0.0094 g and in Asia Pacific - Singapore to 0.055 g, a factor of 5.9 across the 16 regions in the dataset. In US East - N. Virginia electricity is 83% of the SAEF total, against 38% in EU North - Sweden, where hardware, construction and amortised training make up most of it.

A location-based figure uses the average emissions of the grid where the servers run. A market-based figure uses the factor of the electricity the company has contracted, so purchased renewables can lower it. Google's 0.03 g uses a market-based factor of 94 g CO2e per kWh, its 2024 average applied to May 2025. Meta's model card for Llama 3.1 gives training emissions of 420, 2,040 and 8,930 tonnes for the 8B, 70B and 405B models on a location-based basis, and states that the market-based figure is zero because it matches all its electricity with renewables.

SOMA uses location-based grid averages, not renewable contracts. Where a provider publishes a market-based figure, report it beside the location-based one and name both. The distinction is set out in the GHG Protocol tiers guide.

SOMA AI (somaai.earth) estimates one exchange of 1210 tokens. The figures include electricity, hardware, data-centre construction and amortised training, and are computed from the dataset, not typed in. Region names link to the factor pages.

The model class moves the result by a factor of 13 between class A and class C in the same region. Across the table, the region and the class together span a factor of 55, which is why a single "grams per prompt" figure without either is not a usable number.

Report the activity, the factor and the basis rather than a grams-per-prompt headline. A defensible line names the number of tokens or messages, the model class, the region, whether the factor is location-based, the dataset version and the range. The method for placing this in a Scope 3 inventory is in the Scope 3 guide, and the free calculator gives a first estimate without an account. The equations are on the methodology page and the full values in the factor table.

• How much energy does a ChatGPT query use? The energy figures behind these grams
• Is AI bad for the environment? Per prompt and in total
• Why AI footprint estimates differ: published figures, dated. Energy, carbon and water side by side