Carbon footprint of mid-size AI models per query and per token
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.
How much CO₂ does one query to a mid-size model emit?
One query of 1210 tokens to a mid-size model (class B, such as gpt-4o, chatgpt-4o-latest, gpt-4.1) emits about 0.053 g CO₂e on the global-average grid, and a million tokens emit 0.0388 kg CO₂e. Served from EU North - Sweden the same query emits 0.0094 g, and from Asia Pacific - Singapore 0.055 g.
For comparison on the same grid, a class A (small) query emits 0.016 g and a class C (frontier) query emits 0.2 g. Class B uses 0.07128907393577297 kWh of facility electricity per million tokens, and its uncertainty band around the class factor is about 39%.
Class B factors by cloud region
| Region | Electricity kg | Total kg CO₂e | g per query | Water L | mL per query |
|---|---|---|---|---|---|
| Asia Pacific - Japan | 0.03512 | 0.040 | 0.054 | 0.213 | 0.29 |
| Asia Pacific - Singapore | 0.03571 | 0.040 | 0.055 | not available | not available |
| EU Central - Germany | 0.02592 | 0.030 | 0.041 | 0.144 | 0.2 |
| EU North - Finland | 0.00580 | 0.010 | 0.014 | 0.324 | 0.45 |
| EU North - Sweden | 0.00274 | 0.007 | 0.0094 | 0.430 | 0.59 |
| EU South - Spain | 0.01210 | 0.016 | 0.022 | not available | not available |
| EU West - Belgium | 0.00960 | 0.014 | 0.019 | 0.188 | 0.26 |
| EU West - France | 0.00380 | 0.008 | 0.011 | 0.264 | 0.36 |
| EU West - Ireland | 0.02014 | 0.025 | 0.033 | 0.106 | 0.15 |
| EU West - UK (London) | 0.01680 | 0.021 | 0.029 | 0.173 | 0.24 |
| EU-27 average | 0.01682 | 0.021 | 0.029 | not available | not available |
| Global average | 0.03445 | 0.039 | 0.053 | not available | not available |
| Netherlands | 0.01912 | 0.024 | 0.032 | 0.247 | 0.34 |
| US Central - Iowa | 0.03365 | 0.038 | 0.052 | 0.227 | 0.31 |
| US East - N. Virginia | 0.02166 | 0.026 | 0.036 | 0.172 | 0.24 |
| US South - Southeast | 0.03073 | 0.035 | 0.048 | 0.170 | 0.24 |
| US South Central - Texas | 0.02675 | 0.031 | 0.043 | 0.106 | 0.15 |
| US West - Oregon | 0.02307 | 0.027 | 0.038 | 0.682 | 0.94 |
Which region is best for mid-size models?
EU North - Sweden gives the lowest total for class B, 0.00714 kg CO₂e per million tokens and Asia Pacific - Singapore the highest at 0.0401 kg, a 5.6 times spread. The spread on electricity alone is 13 times.
How much water does a mid-size model use per query?
A 1210-token query to a class B model consumes 0.15 mL of water in US South Central - Texas, the lowest-water region with a figure, and 0.94 mL in US West - Oregon, the highest. 14 of 18 regions have a water figure.
How is a class B total split between its four terms?
On the global-average grid, electricity is 88.7% of the class B total, hardware 2.6%, data-centre construction 1.2% and amortised training 7.6%. In EU North - Sweden electricity falls to 38.4% and in Asia Pacific - Singapore it rises to 89%. The conservative high case adds 59% to the class total.
Which models are class B?
The dataset's mapping table assigns 53 model name patterns from 13 providers to class B, and a star stands for any text. 2 of the 36 mapping rows for this class are flagged as reasoning models by default.
- Google gemini-*-pro, gemma-*-27b, gemini-1.0-pro, gemini-1.5-pro, gemini-pro, gemini-omni, gemma-*-31b, gemma-
- Meta llama-3.1-70b, llama-3.3-70b, llama-4-scout, llama3-70b, llama3-1-70b, llama3-3-70b, llama-3-70b, llama4-scout
- Mistral mistral-small, mistral-medium, mixtral-8x22b, mistral-small-4, codestral, devstral, magistral, mixtral-8x
- Alibaba qwen*-plus, qwen3-32b, qwen*-122b-a10b, qwen*-27b, qwen*-32b
- Anthropic claude-sonnet-*, claude-3-sonnet*, claude-3-5-sonnet*, claude-3-7-sonnet*, claude-*-sonnet
- OpenAI gpt-4o, chatgpt-4o-latest, gpt-4.1, gpt-oss-120b
- Amazon nova-pro, nova-*-pro, titan-text-express
- Cohere command-r, aya-expanse-32b, aya-23-35b
- AI21 jamba-mini, jamba-1.5-mini
- DeepSeek deepseek-r1-distill-llama-70b, deepseek-r1-distill-qwen-32b
- NVIDIA nemotron-*-super, nemotron-super
- Perplexity sonar, sonar-pro
- Tencent hunyuan-a13b
- Class A (small)
- Class B (mid-size)
- Class C (frontier)
Where do these numbers come from and how do I cite them?
Figures per query are one plain exchange, before the Tier 2b reasoning uplift. Electricity term. GPU energy per token from the ML.ENERGY Leaderboard v3, a facility overhead of 1.20, and grid carbon intensity from EPA eGRID 2023 and Ember 2023. Embodied hardware, data-centre construction and amortised training terms are documented on the methodology page. Water combines on-site cooling water (Li et al. 2025) with the water embedded in grid electricity (Reig et al. 2020). The modelled range of the models inside a class, around the class value, is −41% to +48% for class A, −39% to +24% for class B and −67% to +55% for class C.
Dataset: AI Inference Emission and Resource Factors for Corporate GHG Inventories (SOMA AI Emission Factors (SAEF)), Llopis-Garcia, G., SOMA AI, 19 October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447, earlier version 2 values at 10.5281/zenodo.22767475). Method: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, arXiv:2606.10660, under peer review.
Partners can fetch any row from the AI factors API, and the whole table is a CSV download.
The full table for every class and region is on the SOMA AI Emission Factors (SAEF) page.