SOMA

    How much energy does a ChatGPT query use?

    For sustainability teams, analysts and journalists who need a defensible energy figure for an AI query. Sources checked on 31 October 2026. Factor version soma-ai-ef-2026.10-dataset-v3.0.0. Last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.

    Part of the SOMA AI Emission Factors cluster.

    How much energy does a ChatGPT query use?

    About 0.3 Wh for a short text query on recent published estimates, and several Wh for long or reasoning queries. Epoch AI modelled 0.3 Wh for GPT-4o with 500 output tokens in February 2025, and Microsoft researchers modelled a median of 0.31 Wh for frontier-scale models on H100 servers. Sam Altman gave 0.34 Wh for the average ChatGPT query in June 2025, without a method. Epoch AI and the Microsoft researchers count inference only, and Altman states no boundary. These are estimates, not meter readings from OpenAI.

    Which figures have been published, and what does each count?

    Read the boundary before the number. Some figures count only the chips, some add servers and cooling, and some are measured on benchmark workloads rather than on a live service. Each row below was opened at its primary source.

    Energy per query, as published. Wh is watt-hours. Training is excluded in every row where the source states a boundary.
    Publisher and dateFigureModel and queryWhat it counts
    Sam Altman, June 20250.34 WhAverage ChatGPT queryNot stated. No method or token count is given.
    Epoch AI, 7 February 20250.3 WhGPT-4o, 500 output tokensModelled. Includes server and data-centre overhead. Epoch calls it relatively pessimistic and gives a range of 0.1 to 4 Wh under other assumptions.
    Oviedo et al. (Microsoft), September 20250.31 Wh (IQR 0.16 to 0.60)Models above 200 billion parameters, 300 median output tokensModelled bottom-up for H100 nodes (8 GPUs, 10 for DeepSeek-R1), with data-centre overhead.
    Jegham et al., May 20250.42 Wh (plus or minus 0.13)GPT-4o, 100 input and 300 output tokensModelled from public API latency and estimated hardware, with overhead. 0.37 Wh without data-centre overhead.
    Google (Elsworth et al.), 21 August 20250.24 WhMedian Gemini Apps text promptActive chips, host CPU and memory, idle machines and data-centre overhead. Google's own service, not ChatGPT.
    Google, 21 August 20250.10 WhSame prompt, narrower methodActive TPU and GPU consumption only. Google calls this an optimistic scenario.
    ML.ENERGY Leaderboard v3.0 (University of Michigan), data as of 16 February 20260.21 WhLlama 3.1 405B, 367 output tokensMeasured GPU energy only, at one operating point of 8 H100 GPUs. Other configurations differ widely.
    MIT Technology Review, 20 May 20250.032 Wh and 1.86 WhLlama 3.1 8B and 405B, per responseMeasured GPU energy of 57 J and 3,353 J, doubled to 114 J and 6,706 J as an assumption for the rest of the system.

    Only the first row names ChatGPT, and it gives no boundary. The others describe other services or open models and are used as proxies. For the full side-by-side with carbon and water, see why AI footprint estimates differ.

    Why does a long prompt use so much more than a short one?

    Because energy grows with the tokens processed. Epoch AI estimates that a 10,000-token input (about a short paper) raises GPT-4o to around 2.5 Wh, and a 100,000-token input (roughly 200 pages) to almost 40 Wh. Jegham et al. estimate 0.443 Wh to 29.078 Wh across 27 models for a long prompt of 10,000 input and 1,500 output tokens, with the top of the range being DeepSeek-R1 on DeepSeek's own servers. Microsoft's test-time-scaling queries, with 5,000 median output tokens, come to a median of 3.91 Wh, about 13 times the short-query figure.

    A single number for "a ChatGPT query" therefore hides a range of more than two orders of magnitude. The useful question is which query, on which model, with how many tokens. The reasoning case has its own page, how much more energy reasoning models use.

    What does SOMA estimate for one exchange?

    SOMA AI (somaai.earth) publishes the SOMA AI Emission Factors (SAEF), an open dataset, and estimates one exchange of 928 input and 282 output tokens, 1210 in all. The figures below are facility-level electricity for that exchange and are the same in every region, because the region changes the carbon and water, not the watt-hours. They are computed from the dataset, not typed in.

    Estimated electricity for one exchange of 1210 tokens, by SAEF model class. Estimates.
    Model classWh per exchangeAgainst 0.3 Wh
    Class A0.03 Wh0.1 times
    Class B0.099 Wh0.3 times
    Class C0.37 Wh1.2 times

    Class B, the mid-size class, is the default for unknown models. Class C covers the largest models, and its estimate sits at 0.37 Wh, close to the upper published values for a standard query. The class choice moves the answer by a factor of 12 between A and C, so it matters more than any other single input. The class mapping and its uncertainty are on the methodology page and in the factor table.

    Is 0.3 Wh a lot?

    Altman's own comparison is that 0.34 Wh is about what an oven uses in a little over one second. For one person that is negligible. The footprint appears at volume. A team sending hundreds of thousands of exchanges a month multiplies the figure by that count, and the class, the model and the share of long prompts then decide the total. Water follows a separate logic and is covered in the water guide.

    What should a company report?

    Report an estimate with its basis, not a single figure presented as a measurement. Name the activity data (tokens, messages or spend), the model class, the region, the factor version and the range. The step-by-step method for Scope 3 is in the Scope 3 guide, and the free calculator gives a first estimate without an account.

    Keep reading