A ChangeEnable Research Project
A disclosure scorecard grading how major AI labs report the energy and water cost of each query. Grades reflect transparency, not environmental performance.
Last reviewed April 2026 · 7 labs tracked
Median text prompt: 0.24 Wh, 0.26 mL water, 0.03 gCO2e. Comprehensive methodology covering TPUs, host CPU/DRAM, idle capacity, datacenter overhead. Excludes image, video, training, networking.
Disclosure gap: No image/video figures. Only on-site water (excludes generation water).
Externally audited LCA with Carbone 4 and ADEME for Mistral Large 2. Reports per 400-token page, not per median prompt — not directly comparable to Google.
Disclosure gap: Different functional unit. No newer model coverage.
Sam Altman blog: 0.34 Wh and 0.000085 gallons water 'average query.' No model specified, no methodology, no audit, no scope. Has not disclosed per-query figures since GPT-3 (2020).
Disclosure gap: GPT-5 and o-series figures absent. Independent estimates put GPT-5 at ~18 Wh and o3 even higher.
Ask OpenAI to disclose →No public per-query energy or water figures for any Claude model. Sustainability reporting at the corporate level only.
Disclosure gap: Everything.
Ask Anthropic to disclose →No per-query inference figures published. Some training-side disclosures in model cards, but no serving footprint.
Disclosure gap: Everything inference-side.
Ask Meta to disclose →No environmental disclosures of any kind for Grok models. Memphis datacenter has drawn separate scrutiny over local air and power impact.
Disclosure gap: Everything.
Ask xAI to disclose →No disclosures. Independent estimates (URI AI Lab) place R1 among the most energy-intensive models tracked.
Disclosure gap: Everything.
What grades mean. A high grade means the lab has published measurable, methodology-backed per-query figures — not that its model is most efficient. See the full methodology for the rubric and weights. Sources: Google Gemini technical paper (Aug 2025), Mistral AI / Carbone 4 / ADEME LCA (July 2025), Sam Altman blog (June 2025), University of Rhode Island AI Lab dashboard, Epoch AI, Shaolei Ren (UC Riverside).