SOMA

    How much more energy do reasoning models use than standard chat?

    For sustainability teams, analysts and engineers who need to know what reasoning modes do to an AI footprint. Sources checked on 31 October 2026. 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 more energy do reasoning models use than standard chat?

    On the AI Energy Score v2 benchmark, turning reasoning on for the same model raised GPU energy per 1,000 queries from 18.35 to 12,791.22 Wh for SmolLM3-3B, which is 697 times. Across the three models tested the factor ran from 154 to 697. That is GPU energy only, on one benchmark. In Microsoft researchers' bottom-up model, queries 15 times longer used 13 times the median energy, 3.91 Wh against 0.31 Wh. The multiplier depends on how many tokens the model writes before it answers. SOMA AI (somaai.earth) treats the multiplier as an assumption in its seat-based estimates, and needs no multiplier when billed tokens are available.

    What do the benchmarks show with reasoning off and on?

    AI Energy Score v2, Hugging Face and partners, 4 December 2025. GPU energy measured with CodeCarbon on the benchmark's custom datasets. The per-query column divides the published per-1,000 figure by 1,000.
    ModelParametersReasoning off, Wh per 1,000 queriesReasoning on, Wh per 1,000 queriesIncreaseWh per query, off then on
    SmolLM3-3B3B18.3512,791.22697 times0.018 then 12.8
    Phi-4-reasoning-plus15B18.429,461.61514 times0.018 then 9.5
    DeepSeek-R1-Distill-Llama-70B70B49.537,626.53154 times0.050 then 7.6

    The source blog post states the increase as between 150 and 700 times for the same model, and says reasoning models use on average 30 times more energy than models with no reasoning capability. The column is labelled GPU energy, which implies that CPU, memory and facility overhead are not in it, and no one should read the factor as an estimate of a company's inventory.

    What happens with longer queries on standard models?

    Two sources on long outputs and long inputs. Both are models of energy per query rather than measurements.
    FigurePublisher and dateWhat it countsRatio to the short query
    0.31 Wh, IQR 0.16 to 0.60, for 300 median output tokensOviedo et al. (Microsoft), Joule 2026, first posted 24 September 2025Models above 200 billion parameters on H100 nodes, 500 input tokens, modelled PUE and batching1
    3.91 Wh, IQR 2.15 to 7.05, for 5,000 median output tokensOviedo et al., same paperSame models, test-time scaling queries 15 times longer13 times
    0.3 Wh, 500 output tokensEpoch AI, 7 February 2025GPT-4o, inference only, relatively pessimistic by Epoch's own description1
    2.5 Wh with 10,000 input tokensEpoch AI, 7 February 2025Same method plus the cost of processing a long input8.3 times
    Almost 40 Wh with 100,000 input tokensEpoch AI, 7 February 2025Same method, roughly 200 pages of inputabout 130 times

    The Microsoft authors call these test-time scaling queries, which is the reasoning case, and do not label them agentic. The long-input rows are not reasoning. They show the other way a single query grows, by reading more. The ratios are our arithmetic from the published values.

    How does SOMA estimate the reasoning uplift for seat-based tools?

    Only where there is no token count. For a seat-licensed tool that exports messages or seats rather than tokens (the Tier 2b case in the Scope 3 guide), SOMA converts each message to an assumed 928 input and 282 output tokens and multiplies the output tokens only by an uplift. The uplift is 1 + s × (k − 1), where s is the share of messages using a reasoning mode and k is how many times more output tokens a reasoning message generates. The share is 5% (range 2.9% to 6.4%), computed from the Anthropic Economic Index release of 27 March 2025, as a share of Claude.ai conversations using thinking. That is a share of conversations and not of tokens, and it may be low for later years. The multiplier k is 25, with a range of 10 to 50, and no public source gives it, so it is the author's assumption and is labelled as one. The central uplift is 2.20.

    Effect of the multiplier on one exchange, holding the share at 5%. Computed from the published defaults.
    Multiplier kUplift on output tokensTokens per exchange
    101.451,336
    25 (central)2.201,547
    503.441,898

    The uplift does not copy the benchmark ratios above, because those are for models run with reasoning on every query, and the Tier 2b share is small. A company that knows its tool is reasoning-heavy should replace the assumption with data, as the next section explains.

    Do billed tokens already include reasoning?

    Yes. As the Scope 3 guide notes, a billed token count already includes the reasoning tokens. For that reason SOMA does not apply the uplift to billed tokens in Tier 2a. Multiplying them would count the reasoning twice. The uplift exists only to approximate those tokens where all the company holds is a seat count or a message count. Where a tool is known to be reasoning-heavy, the better course is to get the billing tokens and use Tier 2a.

    What does the uplift do to a Tier 2b estimate?

    SAEF Tier 2b estimate for 1,000 chat messages served from US East - N. Virginia, including the central reasoning uplift. The band is the dataset's uncertainty band on tokens. Estimates, in kg CO₂e.
    Model classCentralBand lowBand high
    Class A, small0.01650.008250.0379
    Class B, mid-size0.05580.02790.128
    Class C, frontier0.2110.1050.485

    The band covers the token assumptions, not the class, the region or the reasoning multiplier, so declare k separately. The factors themselves are in the full factor table and the equations are on the methodology page.

    What should a company report?

    Report which tier produced each line. For API use, use billed tokens, which carry reasoning already. For seat tools, use the Tier 2b estimate and state the share and multiplier as assumptions with their ranges. Do not quote a benchmark factor such as 154 or 697 as a company's uplift. The free calculator applies both tiers, and the Scope 3 guide sets out how to disclose them. The figures guide shows how published numbers differ.

    Where do the factors come from, and how should they be cited?

    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.

    Keep reading