SOMA

    What are the gCO2e per 1000 tokens emission factors for Claude, Gemini and Llama models?

    For analysts who need an emission factor per 1,000 tokens for a named model. Estimates from the open SAEF dataset, factor version soma-ai-ef-2026.11-dataset-v3.1.0, checked on 12 November 2026.

    Page updated

    What are the gCO₂e per 1,000 tokens emission factors for Claude, Gemini and Llama models?

    SOMA estimates 0.012 g CO₂e per 1,000 billed tokens for a small model, 0.039 g for a mid-size model and 0.15 g for a frontier model at the world-average grid. There is no factor per brand. Claude Haiku, Gemini Flash and the small Llama models sit in class A, Claude Sonnet, Gemini Pro and Llama 70B in class B, and Claude Opus and Llama 405B in class C. Claude and Gemini models are placed by price tier and Llama models by published size, because the closed-model providers do not disclose how large their models are.

    Which class does each model map to, and on what basis?

    g CO₂e per 1,000 billed tokens at the world-average grid, at the dataset's default mix of 85% input and 15% output tokens. The basis is the dataset's rule. Tier means the class is inferred from the vendor's own price tier, and size means the published parameter count. Own factor means the model was benchmarked and uses its own benchmarked energy per token instead of the class mean. Model identifiers are examples, and version numbers change.
    ModelClassBasisWhyCentralHigh case
    Claude HaikuAtierAnthropic's smallest tier0.012 g0.019 g
    Claude SonnetBtierAnthropic's mid tier0.039 g0.062 g
    Claude OpusCtierAnthropic's frontier tier0.15 g0.23 g
    Gemini FlashAtierGoogle's efficiency tier0.012 g0.019 g
    Gemini ProBtierGoogle's mid tier0.039 g0.062 g
    Llama 3.1 8BAown factorPublished size, 8 billion parameters0.0071 g0.011 g
    Llama 3.3 70BBown factorPublished size, 70 billion parameters0.047 g0.075 g
    Llama 4 ScoutBown factorMixture of experts, 109 billion total and 17 billion active0.046 g0.074 g
    Llama 3.1 405BCsizePublished size, 405 billion parameters0.15 g0.23 g

    The central value includes electricity, embodied hardware, data-centre construction and amortised training. The high case raises host uplift, idle capacity with its tied utilisation, and facility overhead to the high end of their ranges. Class pages give the full detail for class A, class B and class C.

    Why is there no figure for Claude, Gemini or Llama as such?

    Because no provider publishes the size of its closed models, and the size, not the name, drives the energy per token. Anthropic and Google do not disclose parameter counts for Claude and Gemini, so SOMA uses the price tier as a proxy. The dataset notes that for some mid-tier models other studies report energy at or above class C, so the assignment is a low-side one. Meta publishes sizes for Llama, so those rows use size, and a mixture-of-experts model such as Llama 4 Scout is classed by its total parameters, because all experts must sit in accelerator memory. Llama models that were benchmarked take their own factor instead of the class mean. A wrong class moves the figure by up to about thirteen times, which is more than the range inside any class. The grid matters as well. These values use the world-average grid, and the same tokens served from a low-carbon region give a fraction of the figure. The factor table has the values for every cloud region and a CSV download, and the methodology page explains the terms.

    Keep reading