Thirsty Tokens

Roughly how much data-center water an AI answer costs.

0mL

Electricity-

GPU time-

Output tokens generated - full cost

The answer it writes. Expensive: one GPU pass per token.

Input tokens read - x0.2

What it reads. Processed in parallel, so far cheaper per token.

Translated

Want your real number, tracked automatically?

The companion tool reads your local Claude Code transcripts, so you never type a token count. CLI, a live desktop widget, a downloadable app, and an MCP server.

claude mcp add thirsty -- python -m thirsty.server

Get it on GitHub →