How reliable are per-prompt AI carbon figures and what are their uncertainty ranges?

Reliable to an order of magnitude, and not to two significant figures. For a mid-size model on the world-average grid, SOMA estimates 0.053 g CO₂e for one 1210-token exchange, with a class range of −39% to +24% and a high case of 0.085 g, about 59% above the central value. The larger risks sit outside those bands. A model placed in the wrong size class moves the number by up to 13 times between the smallest and largest class, and the cloud regions of the dataset differ by about 12 times between the lowest and highest grid for a mid-size model. Published provider figures are self-reported for one service and one boundary, so they cannot be used as an error bar for another service.

The high case moves the three facility uplifts together. The host-server uplift goes from 1.50 to 1.88, idle capacity from 1.10 to 1.20 with utilisation of provisioned capacity tied to it at 0.60 instead of 0.75, and facility overhead (PUE) goes from 1.20 to 1.45. Training is held central. The paper calls this a conservative high case, reported beside the central value and not a confidence interval or a worst case. The central value includes a share of embodied hardware, data-centre construction and amortised training, while the cost of developing the model is shown as a separate line and not added.

• Counting messages instead of tokens. When only messages are known (Tier 2b), the token band is -50% to +130%, taken from the interquartile ranges of measured token counts per exchange.
• Spend instead of use. The spend-based factor of 0.09613 kg CO₂e per euro carries ±300%. In the method paper, spend-based figures for API purchases ran 1.7 to 12.3 times a token-based estimate.
• Training amortisation. Shortening the service life of the model, over which its final training run is spread, from 12 to six months raises the class B factor by about 7.6%, and moving the hardware inputs together changes it by under 2%. These terms are not what drives the width of the range.
• Wrong class. Providers do not disclose model sizes. SAEF assigns closed models by tier, and states that some assignments are low-side. This error is larger than any band in the table and is the first thing to check.
• Region. The same exchange is estimated on grids that differ widely. If the provider does not say where a prompt ran, SOMA applies the world-average grid and says so.

Because they draw different boundaries, not because one is wrong. Google's August 2025 paper by Elsworth et al. reports 0.24 Wh for the median Gemini Apps text prompt, counting idle machines and facility overhead, while its companion blog post gives 0.10 Wh for active chips alone, a factor of 2.4 apart for the same prompt. Epoch AI (7 February 2025) modelled 0.3 Wh for a GPT-4o query with 500 output tokens and around 2.5 Wh with a 10,000-token input. Differences like these come from the boundary, the length of the query and the age of the hardware assumptions. The guide on why estimates differ lines the figures up side by side.

Use it to rank options and to size an inventory line, not to claim a precise footprint for one conversation. State the model class, the region, the token counts and the factor version, and print the range in the same row as the figure. A figure shown as a range is easier to defend to an assurance provider than a precise-looking number without a method. The band definitions are on the methodology page, the class values are in the factor table and the method is in arXiv:2606.10660.

• What explains the range between 0.3 Wh and 3 Wh per AI query? Why energy per query runs from 0.3 to 3 Wh
• What is the carbon footprint of one ChatGPT prompt? Published carbon figures with SOMA estimates by region
• What methodology should a company use to estimate the energy use of generative AI inference? Choosing a method before collecting data