TreadLightlyAI
AI Transparency Card
Models
Certified renewable electricityFree, Starter, and Builder plans
Energy source not disclosedAdvanced plan only
Inference
Training
Inference is only part of the overall footprint. Training is a significant cost too. We have found no primary source for how any of these creators powered training, and we are not aware of any large model trained on renewable electricity.
| Creator | Energy |
|---|---|
| Google DeepMind | Unknown |
| OpenAI | Unknown |
| Anthropic | Unknown |
Compromises we make
Most energy numbers are estimates
Only one model reports its own energy use: Green Large. Until August 2026 the default model reported it too. It now runs on Infomaniak, which reports none, so most of what you see is a token-based estimate.
We made that trade knowingly. Infomaniak has the strongest environmental record of any host we use, and the model it serves passes our safety bar. Neither of those turns an estimate into a measurement, so we label every estimate and say which is which.
Anthropic infrastructure
Two Claude models are available on Advanced. Anthropic has not published where Claude runs or how that electricity is sourced.
Use of Claude Code
Sonnet, Opus, and Fable assist development.
Environmental
Water Use
Energy Methodology
- Reported by the provider
- Green Large: 115 µWh per token, reported by GreenPT
- Estimated by us
- Green Tiny and Green Small: ~108 µWh per token
Claude Haiku 5.5: ~78 µWh per token
Claude Sonnet 5.5: ~240 µWh per token
Green Docs: ~60 µWh per token. It prepares documents you upload for search; this is our roughest estimate. - Overhead
- Estimates include 20% for cooling, networking, and storage.
- Units
- A µWh is a millionth of a watt-hour. Tokens count what you send and what comes back. 1,000 tokens on Green Small is about 0.1 Wh.
Data Center Electricity
Each label comes from the provider's own published statements, linked in the row. We have checked none of it ourselves. Last checked 2026-10-28.
- Supply: hydro and solar, certified naturemade star and TÜV SÜD EE01, through Geneva's utility, SIG (2024).
- PUE (power usage effectiveness): 1.085 across all Infomaniak sites (2024).
- Heat reuse: D4 sends its heat into Geneva's district heating. Energy reuse factor 1.00 on d4project.org, read 2026-10-28.
- Its auditor reviewed the emissions figures, not the certificate purchases (2024).
What each label means: energy terms.
Web SustainabilityB
- We check against
- The W3C Web Sustainability Guidelines, on every public page.
- Tool
- wsg-check
Ethics & Policy
Model Inclusion Standard
- We exclude
- Below C overall or any F on the FLI Safety Index.
Any model creator that dissolved its safety team.
A creator that fails is excluded wherever its model is hosted. - We require
- At least one of: green-hosted, low energy per token, safety-leading, or open-weight.
Safety Scores
Grades from the Future of Life Institute (FLI) AI Safety Index, Summer 2026.
Commitments