Which AI model has the lowest carbon footprint? GPT, Claude, Gemini, Llama, DeepSeek and Mistral compared

Across the 12 models below, one exchange of 1,210 tokens comes to an estimated 0.0097 g CO₂e for llama-3.1-8b and 0.2 g for deepseek-r1, a spread of about 21 times. SOMA AI (somaai.earth) produces these estimates from the open SOMA AI Emission Factors (SAEF), which assign each model a size class and a serving region. In SAEF the class and the region drive the spread. The method cannot see differences between providers inside a class. They are estimates, not measurements by any provider.

Each row is one exchange of 928 input and 282 output tokens. SAEF does not know the size of closed models, or any provider's hardware or serving region. It assigns a size class from the provider's published price tier, or from the parameter count for open models, and assumes a default serving region. That region is an assumption, not a statement of where any provider runs the model.

SOMA classes Claude Sonnet and GPT-4o as mid-size from their published tier, not from a disclosed size. SAEF's own mapping notes add that Jegham et al. report energy that would place both at or above the largest class, so the mid-size assignment is a low-side one. Models that reason by default, such as gpt-5, gpt-5-mini and deepseek-r1, are shown for the same fixed token count, so their rows are a floor. See how much more energy reasoning models use.

Only two of the providers in the table publish a per-query figure we could check. We found no published per-query figure from Anthropic, OpenAI, Meta or DeepSeek in our source check. Sam Altman gave 0.34 Wh for the average ChatGPT query on his personal blog in June 2025, with no stated boundary, so it is not an OpenAI figure.

The two are not comparable with each other or with the table above. Google's 0.24 Wh sits between the SAEF mid-size and largest class estimates for a 1,210-token exchange. Mistral's 1.14 g is about 27 times our estimate of 0.043 g for mistral-large, which is set for an exchange of 928 input and 282 output tokens against Mistral's 400-token response. Mistral counts the marginal impacts of inference, including upstream impacts such as server manufacturing, and excludes users' devices. SAEF also counts hardware manufacturing, data-centre construction and amortised training, on the location-based French grid. The Mistral summary does not name the model behind the figure or break it down, so we cannot say where the difference comes from. See the published-figures guide.

SAEF cannot tell. It gives every model in a class the same factor, so gpt-5 and claude-opus-4 return identical numbers. That is a limit of the method, not a finding that the models are equal. At the same region, class C is 13 times class A. The region is the second lever. The same class C exchange estimates 0.13 g in US East - N. Virginia and 0.037 g in EU North - Sweden, 3.6 times apart, as the country comparison shows in full. The third lever is reasoning, which multiplies the tokens a model generates for the same question.

Report the model class, the region, the token source and the factor version beside each figure. Where you can choose, a smaller class and a lower-carbon region move the estimate more than a change of provider. The Scope 3 method is in the Scope 3 guide, the class mapping is on the methodology page and in the factor table, and the free calculator applies them to your own usage.

• How much more energy do reasoning models use than standard chat? Why the same question can cost many times more
• What is the carbon footprint of Claude, and how does it compare with ChatGPT? Haiku, Sonnet and Opus by region
• How much energy does DeepSeek use, and is it more efficient than ChatGPT? Published figures for DeepSeek-R1 and its distills