Why it matters
The per-prompt energy figures quoted everywhere describe chat, not agents. His logs put one agentic prompt near 150 Wh against 0.24-0.34 Wh for a chat turn, a gap of roughly 600x that reframes most published AI footprint math.
Tokenmaxxing read
About 96% of his tokens were cache reads and only 0.4% were output, so an agent's cost in watts or dollars is mostly context being re-fed. Cache reads run near a tenth of fresh input energy, which makes context discipline the lever for both bills.
Source takeaway
He checks exact session-log token counts against three published methods (Watershed tiers, Simon Couch's coefficients, claude-carbon). His median session, ~600 Wh and ~10 million tokens, dwarfs Couch's published 41 Wh, 592k-token median.
