00The answer
In Turbo Rig's audited week (Sept 14-21, 2026; gate live Sept 16-21), reviewer spend totaled $416.49 for 141 merged PRs: about $2.95 per merged PR, on a mixed basis (gate-window spend divided by full-week merges). 50.9M reviewer tokens ran, 94.7% served from cache. Every figure re-derives from the gate log; the data is downloadable.
01Where the numbers come from
$416.49 is the sum of the cost field across all 868 gate-log rows; 141 is the count of in-window merges across five repositories. Divide one by the other and you get $2.95. The basis caveat matters: gate telemetry covers Sept 16-21, merges cover Sept 14-21, the same mixed basis as the reviews-per-PR ratio (6.16).
The full claim tables, the arithmetic identities and the corrections log are public at /evidence, with machine-readable downloads (CSV · JSON).
02Why it is this cheap
94.7% of reviewer tokens came from cache: the reviewer re-reads mostly-cached context, not fresh input. Deterministic checks (tests, type checks) run first and cost $0; in the audited week, 24 of 868 gate runs were check-lane verdicts that never touched a paid model.
You bring your own subscriptions; the rig is a thin layer that routes work to the model that fits and shows you what each merge cost.