Your AI runs on electricity.Meet the pig that counts it.
One command reads your local agent logs and turns tokens into watt-hours. Nothing leaves your machine.
The global trough
All hogs together have eaten kWh
≈ 1.6 house-days
What does a watt-hour even mean?
Drag the tokens. The pig does the math.
- Tokens (a coding session)
- 50k
- Electricity
- 9.5 Wh
≈ 32 seconds of microwave
assuming a mid-size model; frontier models run 2 to 3 times hotter
Run it. Read it. Brag about it.
$ npx watthog
🐷 watthog · estimated electricity use of your LLMs
9,419 assistant messages · Claude Code, OpenCode, Cursor, Copilot
TOTAL ESTIMATE
Energy 11 kWh range 4.0 - 32 kWh
CO2e 4.6 kg Water 12.6 L
≈ 953 phone charges · 11 dishwasher runs
LAST 14 DAYS
06-07 ████████████████████████ 713 Wh
06-08 ████████████████████ 582 Wh
06-09 ███████████████████████ 697 Wh
06-10 ████ 118 Wh
06-11 ███ 97 Wh
Estimates only. The hog is not a scientist.
Run it
npx watthog scans Claude Code, OpenCode, Codex, Cursor and Copilot logs already on your disk.
Read it
Tokens per model become watt-hours, CO2e and water, always with honest ranges.
Feed the leaderboard
Opt in with watthog submit. Aggregates only; your prompts never leave home.
This week's heaviest hogs
Full trough- 1@e2e_verify_deleteChonkclaude-opus-4-8claude-4.5-opus-high-thinkingclaude-opus-4-7claude-opus-4-7-thinking-highgpt-52.6 kWh≈ 1 hot shower
Put yourself on the board
One opt-in command and you get a share card of your own. Aggregates only, never your prompts or paths.
@ci_gremlin
Unit · rank #4 this week
6.2 kWh
this week
88 kWh
all-time
1.2 kWh
per day
How the pig counts
Providers do not publish per-model energy figures, so Watthog maps every model to a size class with a Wh-per-1k-token factor and always shows a low-to-high range, never a false-precision point value.
small Haiku, mini, flash
0.03 Wh / 1k
medium Sonnet, GPT-4o, ~70B
0.19 Wh / 1k
frontier Opus, GPT-5, o3
0.45 Wh / 1k
Wh per 1,000 output tokens. A frontier model eats roughly 15x what a small one does for the same work, which is why the model mix matters more than the token count.
Factors are anchored in measured benchmarks: AI Energy Score, EcoLogits and Google's Gemini disclosure. Input tokens are weighted at 1/8 of output, cache reads at 1/80. All figures are estimates. The hog is not a scientist.