Top-line summary with sources. Version 0.8, August 2026.
Receipt counts prompt submissions. It records three values per prompt: a timestamp, the site, and whether the prompt ran in a standard or a reasoning ("thinking") mode. It does not read, store or transmit anything typed. Nothing leaves the device.
Counts are multiplied by a published per-prompt energy factor for each class of use. Every figure is reported as a low to high range. There is no single-point number anywhere in the tool.
Evidence grades: A = first-party disclosed with a published method, or directly measured. B = independent published measurement. C = modelled from A or B inputs with stated assumptions. D = self-reported or qualitative.
| Class | Wh per prompt | Grade |
|---|---|---|
| Standard text | 0.24 to 0.34 | A |
| Reasoning / thinking | 7.6 to 33 | B |
| Image generation | 0.3 to 11.5 | B |
0.24 Wh is Google's disclosed median for a Gemini Apps text prompt, from the only first-party measurement published with a full methodology. It accounts for active accelerator power, host system energy, idle machine capacity and datacentre overhead. Google notes that a narrower method counting only active TPU and GPU consumption would give 0.10 Wh, and describes that as an optimistic scenario that substantially underestimates the real footprint.
0.34 Wh is the average ChatGPT query figure stated by Sam Altman. No methodology was published, so it is not directly comparable with Google's median.
Independent corroboration for this order of magnitude comes from Epoch AI, which estimates around 0.3 Wh for a typical GPT-4o query.
7.6 Wh is the reasoning-per-query figure from Hugging Face's AI Energy Score v2.
33 Wh comes from Jegham et al. (2025), an infrastructure-aware benchmark across 30 commercial models. It finds o3 and DeepSeek-R1 consuming over 33 Wh per long prompt, more than 70 times GPT-4.1 nano. Claude 3.7 Sonnet with extended thinking sits lower at around 17 Wh. The same paper puts a short GPT-4o query at 0.42 Wh, which is a useful internal check against the standard-text range above.
Peer-reviewed corroboration: a 2026 Joule paper finds reasoning queries of roughly 5,000 output tokens raise energy use about 13 times versus standard queries, for the same model.
Note the caveat: the higher figures are for long prompts. A short reasoning query will sit well below 33 Wh, which is why the range is wide and why the tool never reports a point value.
2.9 Wh is the mean per image from Luccioni, Jernite and Strubell, the first systematic measurement of inference energy across 88 models and ten tasks. 11.5 Wh is Stable Diffusion XL, the least efficient model they tested, which the authors describe as close to a full phone charge per generation.
0.3 Wh is the low bound, from a 2025 study generating 512x512 images with a smaller open model, roughly ten times less than Luccioni et al. reported. The difference is largely resolution, parameter count and denoising steps.
All of these measure open models. Proprietary image generators publish nothing, and Google's Gemini figures explicitly exclude image and video prompts.
The low bound for standard prompts uses Google's disclosed 0.03 gCO2e, which reflects their clean energy procurement. Everything else is energy multiplied by a UK average grid intensity of roughly 200 gCO2e/kWh. No provider discloses which region served a given request, and carbon intensity varies several-fold by region and time of day, so this is a Grade C figure by construction.
Published per-prompt water estimates differ by around 170 times. Google reports 0.26 ml for a median text prompt; Mistral's life cycle assessment puts a 400-token response at 45 ml. They are not measuring the same thing:
An earlier version of Receipt derived thinking and image water by scaling energy at a fixed ratio. That inherited an on-site-only scope and assumed constant cooling design worldwide, so it was removed. The tool now explains the absence rather than showing a number it cannot defend.
Background on water scoping: Li, Yang, Islam and Ren, "Making AI Less Thirsty", Communications of the ACM 68(7).
The tool states these exclusions in the interface, not only in documentation.
Reasoning-mode detection relies on artefacts appearing in the response, which is more robust than reading the interface but will break when vendors redesign. A manual override is provided. Image generation cannot be detected reliably and is logged by hand.
The reasoning and image ranges are wide because the underlying literature is genuinely uncertain, not because the tool is being cautious. Anyone quoting a narrow figure for either is overstating what is known.
Factors are reviewed quarterly and carry a review date in the code. If the figures are out of date, the tool is out of date, and it says so.
The reasoning and image ranges were widened after checking them against the primary literature rather than secondary summaries. Reasoning moved from 10 to 40 Wh to 7.6 to 33 Wh; image generation moved from 1 to 3 Wh to 0.3 to 11.5 Wh, the earlier range having been too narrow at both ends. The in-product prompt that flags reasoning mode now says "often 20 to 60 times" rather than a flat "50 times".
If a figure here is out of date or wrong, that is the most useful thing you can tell me: marie@springcorp.co.uk
Back to Receipt · Privacy
Receipt is an independent project. It is not selling you an AI product and never will.