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

For sustainability teams and analysts who need a number for the Gemini their company uses. Sources checked on 20 November 2026. 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

Part of the SOMA AI Emission Factors cluster.

What has Google published about the carbon footprint of a Gemini prompt?

Google reports 0.03 g CO₂e, 0.24 Wh and 0.26 mL of water for a median Gemini Apps text prompt. That is Google's own figure, from its August 2025 paper, and it is the only per-query figure we could check from Google, OpenAI or Anthropic. It covers a median text prompt in the Gemini Apps, not every Gemini model, image or video request, and not training. Google counts the whole serving stack, including idle machines and data-centre overhead, and reports carbon on a market-based basis. The 0.26 mL counts water consumed at the data centre, at a fleet water usage effectiveness of 1.15 L per kWh.

Google also gives 0.1 Wh for the same prompt on a narrower, accelerator-only method. The gap between the two is the boundary debate, set out on why Google's Gemini figure differs from academic estimates.

How do Gemini Flash and Gemini Pro compare in different regions?

Google's figure is a single median, so it cannot say what a Flash or a Pro request costs, or where it was served. SOMA AI (somaai.earth) fills that gap with estimates from the open SOMA AI Emission Factors (SAEF). SAEF assigns Flash to its smallest class and Pro to its mid class from Google's published tiers, not from a disclosed size. The regions below are Google Cloud regions, and the choice of region is yours when you call Gemini through Vertex AI.

SOMA AI estimates (SAEF) of g CO₂e for one exchange of 1,210 tokens. Energy is the same in every region.
ModelSAEF classWhus-east4 (Ashburn, Virginia)europe-north1 (Hamina, Finland)europe-west1 (St. Ghislain, Belgium)
Gemini Flash, efficiency tierClass A0.030.01 g0.0037 g0.0054 g
Gemini Pro, mid tierClass B0.0990.036 g0.014 g0.019 g

Pro is 3.4 times Flash in the same region, and the same Pro exchange is 2.6 times lower in europe-north1 than in us-east4. These estimates are for an exchange of 928 input and 282 output tokens, a different unit from Google's median prompt, so do not read the two as a pair.

How does Gemini compare with ChatGPT and Claude?

SAEF gives models of the same class and the same region the same number. A real difference between providers, if there is one, is not visible in this method. Google's published figure is on Google's own boundary and cannot be set against an estimate built on another one. Sam Altman's 0.34 Wh for the average ChatGPT query, from his personal blog with no stated boundary, is not an OpenAI figure.

Estimated g CO₂e for one exchange of 1,210 tokens in us-east-1, by tier, for comparison only. Estimates.
SAEF classClaudeOpenAIGoogle Gemini
Class Aclaude-haiku-4, 0.01 ggpt-5-mini, 0.01 ggemini-2.5-flash, 0.01 g
Class Bclaude-sonnet-4, 0.036 ggpt-4o, 0.036 ggemini-2.5-pro, 0.036 g

The Claude side is on the Claude page, and twelve models side by side are on the model comparison.

What should a company report for Gemini?

Take token counts per model from Google Vertex AI through Cloud Monitoring and the billing export, and use the endpoint location as the serving region. For seat-licensed Gemini for Workspace, use seats and activity. Report the class, the region, the factor version and the range, and say that Google's own per-prompt figure covers a median Gemini Apps text prompt on its boundary. SOMA AI estimates this use and does not measure it. The method is in the Scope 3 guide and on the methodology page, and the free calculator gives a first estimate without an account.

Short answers

How much energy does a Gemini prompt use?
Google reports 0.24 Wh and 0.03 g CO2e for a median Gemini Apps text prompt, counting the whole serving stack and excluding training. SOMA AI estimates, from its open SAEF emission factors, 0.03 Wh for a Flash-tier exchange of 1,210 tokens and 0.099 Wh for a Pro-tier one. The two are not the same prompt, so they are not a like-for-like pair.
Is Google's 0.03 g figure the carbon footprint of Gemini?
Only of a median Gemini Apps text prompt on Google's boundary, which uses market-based carbon and excludes training. Google also gives a narrower 0.1 Wh for the same prompt on an accelerator-only method. Other methods give different numbers for the same service.
Is Gemini lower or higher than ChatGPT and Claude?
SAEF cannot tell them apart inside a size class, because it assigns each model a class from its provider's tier. Google is the only one of the three with a published per-query figure we could check, and that figure is on Google's own boundary, so it does not rank Gemini against the others.
Can I estimate Gemini use through Vertex AI by region?
Yes. Vertex AI exports token counts through Cloud Monitoring and the billing export, and the endpoint location is the serving region. SOMA AI estimates, it does not measure, so the result is a class and region estimate with a range.

Keep reading