Cost ledger
In plain terms:the live-drafting path is budget-guarded, so a run can't quietly exceed the cap. Spend is computed from real reported tokens against a pinned price list, and a fail-closed hard stop blocks any call that would cross it.
$0.00
spent this run
$5.00
hard cap (fail-closed)
0
live calls
REPLAY
serve mode
Deterministic stub drafting — no live Gemini calls, $0.00 spend. A real Gemini run is recorded (see Eval); the public demo stays REPLAY-only.
How the cap holds
- Cost = real API-reported tokens × a pinned, versioned price table (not an estimate).
- Before every live call, a fail-closed guard blocks it if spent + next-estimate would exceed the cap.
- A batch threads cumulative spend, so the cap holds across the whole run — not just per call.
- An unknown model id fails loud (never silently prices at $0); a billed-then-failed call still records its cost.
- The price table was pinned + verified against official Gemini pricing for the recorded run; it must be re-checked before any future live run (never trusted from memory).
Pinned price table (2026-06-18)
| Model | Input $/1M | Output $/1M |
|---|---|---|
| gemini-2.5-flash | $0.3 | $2.5 |
| gemini-2.5-flash-lite | $0.1 | $0.4 |
| gemini-2.5-pro | $1.25 | $10 |