Estimates changelog
The coefficients file and the source registry share a single version. Every change to a coefficient, a source, or a reviewed constant lands here, dated.
v1.3.12026-09
Coefficients, sources and derived constants (including the home-page ticker rate) are reviewed quarterly. Current review stamp: 2026-09.
v1.3.1
2026-09-26current- Corrected Google’s water figure: 6.1B gallons was its data centers’ consumption in 2023. For 2024 Google reported ~7.7B gallons for its data centers (~8.1B company-wide). Council Bluffs, Iowa is now ~1B gallons (2024), and the ~10,000-gallon Texas site is Pflugerville.
- Corrected the ~519 mL per 100-word email figure: it comes from a 2024 Washington Post analysis with UC Riverside and counts cooling plus electricity water (our Mid boundary), not the full lifecycle. The High scenario now rests on Mistral’s LCA alone.
- Clarified Mistral’s LCA: it disclosed water and carbon per response but no energy figure, so the ~1.5 Wh behind the High factor (30 L/kWh) is a site assumption. Its 281,000 m³ covers Mistral Large 2’s training plus its first 18 months of use.
- Low scenario: now cites Microsoft’s reported fleet WUE (~0.30 L/kWh) and discloses that Google’s and OpenAI’s own per-prompt figures imply ~1 L/kWh, above the Low factor.
- Relabeled equivalents and factors: the hamburger is a 150 g patty (~2,500 L); the transatlantic flight (~700 kg CO₂e) is a midpoint between fuel-only and full-warming calculators; the tree (~21 kg CO₂/yr) is a commonly cited round figure, not a USDA value; the toilet flush is the 1.6 gal federal maximum; the carbon Mid factor (350 g/kWh) is a rounded value slightly below recent eGRID US averages.
- Ticker (~230 L/second) reviewed: its ~0.8 Wh per query is now disclosed as a blended site assumption between standard- and frontier-class exchanges.
- Registry: Jegham et al. is now tiered as a preprint; added sources for the Washington Post analysis, Microsoft’s water efficiency and Meta’s Talavera de la Reina campus; removed an untraceable USDA attribution.
- Company pages: fixed the OpenAI heavy-use example (100 prompts a day for a year is ~68 gallons at Mid, not ~25) and replaced a Chile data center story wrongly credited to Meta (the Cerrillos project was Google’s) with Meta’s Talavera de la Reina campus.
- No coefficient values changed in this release.
v1.3.0
2026-06-10- Introduced the central source registry: every coefficient, equivalent and dataset entry now references registry IDs instead of free-text source strings, and every number rendered on the site carries a citation marker resolving to it.
- Added structured references (authors, venue, year, access date) and evidence tiers (official disclosure, peer-reviewed, government, industry report, news, site assumption) for all sources.
- Added citations for everyday equivalents previously sourced inline: IEA streaming energy, EPA vehicle figures, USDA tree sequestration, ICAO flight emissions, World Aquatics pool volume, plus a news-tier entry for per-site local reporting (Memphis, West Des Moines, Santiago).
- The global ticker constant (~230 L/second) is now a quarterly-reviewed value; its derivation and last review date are shown wherever it appears.
- Added a pendingModels list for models that exist but have no published energy data (currently Claude Fable 5): they appear labeled 'no data yet' and produce no estimate.
- No coefficient values changed in this release.
v1.2.0
2026-06-10- Video generation band re-anchored to published measurements: 25 / 150 / 1,000 Wh per 5-second clip (MIT Technology Review 2025 measured ~940 Wh for a current-quality open model; Sora/Veo-class providers disclose nothing).
- Image generation re-anchored to Luccioni, Jernite & Strubell (FAccT 2024), mean ~2.9 Wh per image; the previous compilation source was retired.
- GPT mini tier reclassified from efficient to standard per Jegham et al. v6 (GPT-4.1 mini ~1.13 Wh/1K tokens, GPT-4o mini ~1.44 Wh/1K tokens); a separate nano entry was added to the efficient class.
- Benchmark anchors refreshed to Jegham et al. 2025 v6 (Nov 2025) throughout.
v1.0.0
2026-06-08- Initial public coefficient set: energy intensity bands for six model classes, water factors for the three accounting scenarios (Low 0.3 / Mid 5.0 / High 30.0 L per kWh), grid carbon scenarios, and the real-world equivalents catalog.
- Gemini received the only model-level override, from Google's official 2025 per-prompt disclosure.
Your number
What does your AI actually use?
Pick your model, set your usage, get the number, with sources.