How tokens and cost are counted

Tokens are what the agent reported. Each turn records input, output and, where the agent provides them, cache reads and cache writes. A session’s Tokens panel splits usage per model into fresh (input and output) and cached (reads and writes). Imported sessions carry whatever the transcript recorded; Antigravity imports, for one, carry none

Cost is computed when you look at it, never stored: tokens multiplied by the model’s price. Prices sync from models.dev; Administration → AI & models → AI Providers → Model pricing shows when the last sync succeeded and warns after 48 hours without one. A model with no known price shows no cost, which means unknown, not free

Active time on the dashboard adds up session durations, so two parallel sessions add to more than the clock. The dashboard’s last seven days and Insights’ thirty-day tiles are different windows

Evaluation runs record the judge’s own token use and cost on the run; flow runs show each participant’s estimated cost; hosted Grok Build runs report cost per the vendor’s own accounting