SOMA

    AI carbon calculator that uses real model and region data

    For analysts and developers looking for an AI carbon calculator that is not a single flat factor. Written by SOMA AI, which publishes one of the options below. All figures are estimates. Factor version soma-ai-ef-2026.11-dataset-v3.1.0. Factors last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.

    Page updated

    Is there an AI carbon calculator that uses real model and region data?

    Yes, with a caveat on the word real. No major provider publishes per-model energy for its commercial models, so any calculator estimates. What can be real is the input data. The SOMA AI free calculator maps the model you name to one of three size classes, applies the electricity grid of the cloud region or country you pick, and splits input from output tokens. It runs on SOMA AI Emission Factors (SAEF), which covers 143 cloud regions on AWS, Azure and Google Cloud and 240 grids. The same method is available in code through the saef Python package.

    What should a calculator take as input?

    Three inputs that decide the result, with the size of each effect in the SOMA dataset. Estimates.
    InputWhy it mattersEffect in the SOMA dataset
    Model classProviders do not publish parameter counts, so models are grouped by tier and benchmark behaviour into small, mid-size and frontier classes.Per million tokens at the world-average grid, class A is 0.0115 kg CO₂e and class C is 0.147 kg, a factor of 13.
    Region and gridThe same tokens use the electricity of the data centre that serves them, so the grid carbon intensity multiplies the electricity term.The grid effect is worked through for Germany, France and the US on the grid comparison page.
    Token splitOutput tokens cost more energy than input tokens, and reasoning adds hidden output tokens.For class B at the world-average grid, one million tokens with 90% input is 0.0357 kg and with 10% input is 0.0854 kg.

    A fourth input is the lifecycle boundary. SOMA counts inference electricity with facility overhead, a share of server hardware, data-centre construction and amortised training. Calculators that count only chip electricity report lower numbers, which is why the boundary should be printed next to the result. Every class is listed under the factor table, which also lists the model names mapped to each of the three classes.

    Which options exist, and what does each cover?

    Neutral summary from each publisher's documentation, read on 12 November 2026. Numbers from any of them are estimates.
    OptionWhat it estimatesModel and region handlingAccess
    SOMA AI calculatorCO₂e, electricity and water of a company's AI use per tool, from seats, messages, tokens or spendModel name to size class, then 143 cloud regions or any country gridFree, no sign-up
    saef packageThe same method applied to your logged callsSame dataset, 143 cloud regions and 240 gridspip install saef, open source
    EcoLogitsEnvironmental impact of generative AI API calls at inference timeWraps provider SDKs. Its methodology is publishedFree, open source
    CodeCarbonElectricity of code running on your own CPU, GPU and RAM, times regional grid intensityMeasures local hardware. Its README sends remote API users to EcoLogitsFree, MIT

    The practical split is where the model runs. CodeCarbon measures hardware you control, EcoLogits estimates calls to someone else's API, and the SOMA calculator and package estimate a whole organisation's purchased AI, including chat seats that never pass through your code. More comparisons, including cloud provider tools and consultancies, are in the tools comparison.

    How reliable is the output?

    Within a class the dataset reports a range, and a wrong class is the largest single error in the method. The grid is a location-based annual average, not the provider's renewable contracts. The equations, the sources and the uncertainty bands are on the methodology page, and the method paper was accepted for publication in Sustainable Production and Consumption in November 2026.

    Keep reading