SOMA

    What is the carbon emission factor per query for GPT-4o in gCO2e?

    For analysts who need a figure for GPT-4o and the basis behind it. SOMA figures are estimates from the open SAEF dataset, factor version soma-ai-ef-2026.11-dataset-v3.1.0, checked on 12 November 2026. OpenAI has not published one.

    Page updated

    What is the carbon emission factor per query for GPT-4o in gCO₂e?

    SOMA estimates 0.053 g CO₂e for one 1210-token GPT-4o exchange (928 tokens in, 282 out) on the world-average grid, with a high case of 0.085 g. That is an estimate and not an OpenAI figure. OpenAI does not publish the size of GPT-4o or its energy per query, so SOMA places GPT-4o in class B (mid-size) by its price tier and applies the class B factor. The result changes with the grid that serves the request, from 0.0094 g in EU North - Sweden to 0.036 g in US East - N. Virginia.

    What do the factors look like by region?

    g CO₂e for one exchange of 1210 tokens on a class B model. Central values include electricity, embodied hardware, data-centre construction and amortised training. The high case sets host uplift, idle capacity with its tied utilisation, and facility overhead at the high end of their ranges. Energy at the world-average grid is about 0.099 Wh.
    GridCentralHigh case
    World average0.053 g0.085 g
    EU North - Sweden0.0094 g0.012 g
    US East - N. Virginia0.036 g0.056 g

    Why class B, and how sure is that?

    The dataset records GPT-4o as a mid-price tier model with an undisclosed size, so class B is a tier-based assignment. It also notes that the energy reported by Jegham et al. would place GPT-4o at or above class C, so B is a low-side choice. GPT-4o mini maps to class A as an efficiency tier. If you consider GPT-4o to be larger, the class C factor on the class C page applies and is about 3.8 times higher. Published energy figures for GPT-4o are 0.42 Wh (±0.13) for 100 input and 300 output tokens from Jegham et al. (May 2025) and 0.3 Wh with 500 output tokens from Epoch AI (7 February 2025). Both are modelled and not measured by OpenAI, and they use different boundaries, which the comparison guide sets out.

    For a company inventory, multiply your billed tokens by the factor for the region that serves you instead of using a per-query figure, and state the class, the region and the version. The factor table and method are open under CC BY 4.0.

    Keep reading