Expenses

OAuth/subscription turns only — orchestrator rows plus chat-app authMode==="oauth" turns; API-key turns are excluded. Weighted tokens, not raw: input ×1, cache-write ×1.25, cache-read ×0.1, output ×5 (API price ratios). Raw counts are dominated by cache reads — every agent turn re-reads its full ~100k+ cached context, so raw totals overstate real burn ~10× (this chart once showed "768k tokens" for an hour that was ~92% cache reads). Kept in lockstep with rowWeightedTokens in the orchestrator's spend_guard.js. Lower bound: the account-wide 5-hour limit also counts interactive Claude Code sessions and other OAuth usage we don't log here, so these lines understate true consumption. The dashed threshold is an editable estimate — Anthropic doesn't publish an exact token/5hr cap and it varies by plan; retune OAUTH_5H_TOKEN_LIMIT at the top of the page script.

Claude Code OAuth/subscription rows only — source==="orchestrator" or authMode==="oauth" from all cost-log rows surfaced by /chat/api/cost-log. This is a true trailing 7-day usage view; it is not the 5-hour/session cap below. Tokens leave the plotted total only when their own logged timestamp ages out of the rolling 7-day window. Raw token total: input + cache-write + cache-read + output.