Carbon footprint of small AI models per query and per token

One query of 1210 tokens to a small model (class A: gpt-4o-mini, gpt-4.1-mini, gpt-4.1-nano) emits about 0.016 g CO₂e on the global-average grid, counting electricity, the share of server hardware, data-centre construction and amortised training; a million tokens emit 0.0115 kg CO₂e. Served from EU North - Sweden the same query emits 0.0025 g; from Asia Pacific - Singapore, 0.016 g.

For comparison on the same grid, a class B (mid-size) query emits 0.053 g and a class C (frontier) query emits 0.2 g. Class A uses 0.021769339805825245 kWh of facility electricity per million tokens, including the facility overhead (PUE) of 1.20.

EU North - Sweden gives the lowest total for class A, 0.00185 kg CO₂e per million tokens; Asia Pacific - Singapore the highest, 0.0119 kg, a 6.4 times spread. In low-carbon regions the hardware and training terms become the larger share of the total, which is why the spread on the total is smaller than the spread on electricity alone (13 times).

• Asia Pacific - Japan
• Asia Pacific - Singapore
• EU Central - Germany
• EU North - Finland
• EU North - Sweden
• EU West - Belgium
• EU West - France
• EU West - Ireland
• EU West - UK (London)
• EU-27 average
• Global average
• Netherlands
• US Central - Iowa
• US East - N. Virginia
• US South - Southeast
• US South Central - Texas
• US West - Oregon

A 1210-token query to a class A model consumes about not available mL of water on the global average, from 0.045 mL in the lowest-water region to 0.29 mL in the highest. On-site cooling water plus the water embedded in grid electricity.

Class A covers gpt-4o-mini, gpt-4.1-mini, gpt-4.1-nano. Providers do not disclose parameter counts, so the class comes from the published tier and benchmark behaviour, and SOMA keeps the mapping current. Reasoning modes (o1, o3, DeepSeek-R1, extended thinking) generate more tokens per answer; when token counts come from billing that is already in the number. No public source gives a reasoning-token multiplier, so for estimates from message counts SOMA multiplies the estimated output tokens by 1 + s_r(k - 1), with k an author assumption and s_r the 5% of Claude.ai conversations in which the Anthropic Economic Index reports thinking switched on (a share of conversations, not of tokens, and likely low for later years).

The conservative high case adds 61% to the class total.

• Class A (small)
• Class B (mid-size)
• Class C (frontier)

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, Llopis, G., SOMA AI, October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447). 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.