the bridges already knew and threw it away — the usage block was on the wire. it is read per round now (a turn that called a tool paid for two responses), turned into watt-hours from the model's active parameters, and parked on the footer's gauge row beside the viewer count, where it is the first thing a narrow window gives up. `/energy` prints the working. it is an estimate and says so everywhere: nobody publishes per-token energy, so the slope is anchored to the only figures anyone has released (google's 0.24 Wh median prompt, epoch's ~0.3 Wh) and the ratios — which rest on announced active-parameter counts — are the half that can be defended. `/energy` leads with them: the same conversation priced across every model, cheapest first, with the closed ones saying their size is a guess. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>