loading coefficients…
Estimate
Energy, water, carbon. Envelope, not a confidence interval. We do not know the building.
Assumed place
Hypothetical. Sets climate for water. Carbon stays the 345 g/kWh fleet proxy unless a staged live CI exists. API at best says us/global.
If diesel
MODELED overlay — same kWh at ~780 g/kWh direct. Not written into the session number. We did not observe this site.
Hour —×
Stylized CAISO-shaped marginal CI, not your live grid. Click the curve.
Tier / savings
Receipts
- 0.24 Wh · 0.26 mL · 0.03 gCO2e — Google, median Gemini Apps text prompt, full-stack (arXiv:2508.15734). Carbon is market-based — not comparable to this tool’s location-based default.
- 1.14 gCO2e · 45 mL / 400-token page — Mistral Large 2, ADEME + Carbone 4 LCA. Wider scope: includes amortized training.
- ~0.3–0.34 Wh / typical query — Epoch AI and Altman’s public figure. A sanity band, not a coefficient.
| footprint | EcoLogits | |
|---|---|---|
| Tokens priced | in + cache-read + out | output only (output_token_count) |
| Range | scenario envelope on screen | min/max; many UIs show the midpoint |
| Grid | site you declare, or live EM / WattTime on the CLI | country average or WOR; no facility |
| Runtime | stdlib, offline statusline | hosted /v1beta/estimations |
| Water | WUE on IT kWh + EWIF on facility kWh | WCF, usage phase only |
| Serving node | not disclosed (optional inference_geo us/global) | not disclosed — authors say so |
CodeCarbon measures watts on your machine, not the API hall. AMD’s EPYC TCO tools compare servers you buy — annual country kgCO2e/kWh × estimated kWh — not where a prompt ran.
Install
Python 3, stdlib only. Statusline never networks. Live grid: FOOTPRINT_SITE + Electricity Maps token, then refresh.
git clone https://github.com/rdsciv/footprint && cd footprint
pip install .
# statusline — .claude/settings.json
{ "statusLine": { "type": "command", "command": "python3 /path/to/footprint/footprint_statusline.py" } }
python3 -m modelfootprint report
python3 -m modelfootprint whatif opus 500kOptional live grid: FOOTPRINT_SITE=virginia and a free Electricity Maps token. Deep dives live in the repo.
| doc | one-line takeaway |
|---|---|
| 01 · Energy of inference | the honest unit is the token, priced by class, with a range |
| 02 · The duck curve | >10× carbon swing by hour of day on solar-heavy grids |
| 03 · Fuel mix & diesel backup | diesel genset ~780 g direct; we cannot see a campus flip |
| 04 · Water | the power plant usually evaporates more water than the datacenter |
| 05 · Which model, when | one tier down + batch API is cheapest and cleanest |