What is the carbon footprint of Claude, and how does it compare with ChatGPT?

About 0.036 g CO₂e and 0.099 Wh for one Claude Sonnet exchange of 1,210 tokens, with 0.01 g for Haiku and 0.13 g for Opus, on SOMA AI (somaai.earth) estimates for the US East - N. Virginia region. We found no published per-query figure from Anthropic in our source check, so these are estimates from the open SOMA AI Emission Factors (SAEF), not Anthropic measurements. At the same size class and region, Claude and ChatGPT models get the same estimate.

SOMA classes Claude Sonnet as mid-size from its published tier, not from a disclosed size, and does the same for Haiku (small) and Opus (frontier). The default region, us-east-1, is an assumption in SAEF, not a statement of where Anthropic serves a request. If your usage runs through a cloud platform, use the region in the usage export.

Opus is 13 times Haiku in the same region, and the same Opus exchange is 3.6 times lower in EU North - Sweden than in US East - N. Virginia. SAEF's mapping notes add that Jegham et al. report energy that would place Sonnet at or above the largest class, so the mid-size assignment is a low-side one.

Between 0.045 and 2.2 mL in the table below. Water counts on-site cooling plus the water used to generate the electricity, so it does not follow carbon. EU North - Sweden has the lowest carbon here and the highest water. The three kinds of water are in the water guide.

SAEF cannot tell them apart inside a class, because it assigns each model a class from its provider's tier. Same class and same region give the same number. A real difference between providers, if there is one, is not visible in this method. Sam Altman's 0.34 Wh for the average ChatGPT query, from his personal blog with no stated boundary, is not an OpenAI figure and is not on the same basis.

Gemini is shown at the same assumed region for comparison only. Google publishes its own figure, 0.03 g CO₂e and 0.24 Wh for a median Gemini Apps text prompt, reported by Google, and we found no equivalent from Anthropic.

Claude for Work reports seats and activity, so SAEF uses its seat-based Tier 2b estimate. On a Sonnet-class model, one active seat at the default of 30 messages a month comes to about 20 g CO₂e a year, and a heavy user at 325 a month to 218 g. One hundred default seats come to 2 kg. These message rates come from the UK Department for Business and Trade's Microsoft 365 Copilot pilot and an OpenAI workspace definition, not from Claude data.

The estimate multiplies estimated output tokens by 2.20 for reasoning. That uses the Anthropic Economic Index share of 5% of Claude.ai conversations using the Thinking feature in its March 2025 release, which is a share of conversations and not of tokens, together with an author-assumed multiplier of 25 for which we found no public source. SAEF applies that default to every chat tool, so it is a placeholder and not a measured Claude figure.

An agentic coding request sends tens of thousands of tokens, mostly cached, so it carries about 70 times the tokens of a chat exchange, and the estimate is more than ten times higher. The figures and the cache effect are on the coding assistants page.

Take token counts per model from the Anthropic console usage view where you use the API, and seats and activity from Claude for Work. Report the class, the region, the factor version and the range, and say that we found no published per-query figure from Anthropic. The method is in the Scope 3 guide and on the methodology page, and the free calculator gives a first estimate without an account.

• Which AI model has the lowest carbon footprint? GPT, Claude, Gemini, Llama, DeepSeek and Mistral compared. Twelve models side by side
• What is the carbon footprint of AI coding assistants like Claude Code, GitHub Copilot and Cursor? Claude Code, Copilot, Cursor and Codex
• What is the carbon footprint of one ChatGPT prompt? Published per-prompt figures and their basis