SOMA

    What is the carbon footprint of AI coding assistants like Claude Code, GitHub Copilot and Cursor?

    For engineering and sustainability teams who need a number for AI coding tools. 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.

    What is the carbon footprint of AI coding assistants like Claude Code, GitHub Copilot and Cursor?

    On SOMA AI (somaai.earth) estimates for a mid-size model in US East - N. Virginia, one inline completion comes to about 0.04 g CO₂e and one agentic coding request to about 0.65 g, roughly 16 times more. A chat exchange of 1,210 tokens sits between them. In SAEF the gap comes from tokens. The method has no term for the tool brand. These are estimates from the open SOMA AI Emission Factors (SAEF) and the tool's real model, region and token count will differ.

    How many tokens does a completion use, and how many does an agent request use?

    SAEF's Tier 2b defaults hold one row for each. An inline completion sends the code around the cursor and returns a few tokens, with a central 1,952 tokens in total and 8 output tokens (interquartile range 3 to 18), from the 2024 Azure OpenAI coding trace of about 16.8 million requests. An agentic request carries a long working context, with a central 85,000 tokens and 400 output tokens, from the OpenRouter apps ranking for the month to 1 October 2026. In that source Claude Code averages 96,884 tokens a request and Codex 85,269, and Anthropic's own documented usage sample gives 71,179.

    Estimated cost of one request on a model of each SAEF class in US East - N. Virginia. Estimates. Includes SAEF's default reasoning uplift on output tokens.
    Model classOne completion, g CO₂eOne completion, WhOne agent request, g CO₂eOne agent request, Wh
    Class A0.0120.0320.170.38
    Class B0.040.110.651.2
    Class C0.150.392.64.6

    Why does cached input lower the estimate?

    Most of an agent request repeats context the provider already holds. The documented Claude Code sample reports 91% of input tokens read from cache, and SAEF uses 91% as its default. SAEF counts a cached token at a small fraction of an uncached one, so the 85,000 billed tokens reduce to about 5,236 effective tokens. Without the cache credit, the same request would estimate 2.7 times higher on a mid-size class. If your usage export reports cached tokens, enter them separately. If it does not, the estimate takes no credit and runs high.

    What does this mean per 1,000 requests?

    We do not have a public count of how many completions or agent requests a developer makes in a year, and we do not invent one. The defaults table has no row for it, so the useful unit is 1,000 requests. Multiply by the count in your own usage data.

    Estimated kg CO₂e per 1,000 requests in US East - N. Virginia, by model class. Illustrative scale only, not a usage statistic.
    Model class1,000 completions1,000 agent requests
    Class A0.012 kg0.17 kg
    Class B0.04 kg0.65 kg
    Class C0.15 kg2.6 kg

    A developer's yearly figure is their request count times these per-request values, so heavy agent use dominates and completions barely register. The interquartile range on tokens is wide, from -51% to +73% for completions on the dataset's own measurements.

    Where does a company find the numbers for each tool?

    It depends on how the tool is billed. GitHub Copilot has no token export, and the admin centre gives assigned seats and active users, which is Tier 2b data. Where Claude Code or Codex runs on an API account, the Anthropic console usage view and the OpenAI usage page export tokens per model, which is better data and uses Tier 2a. For any tool that shows only seats, use the seat-based estimate and say so.

    What should a company report?

    Report coding tools as their own line, split between completions and agent use, with the token source, the model class, the region and the factor version. Do not carry a chat default of 1,210 tokens over to coding. The method is in the Scope 3 guide and on the methodology page, and the free calculator accepts tokens or seats.

    Keep reading