What you actually pay
Authorized max, metered actuals, automatic refunds, the second-score premium, and the 7% fee, with real numbers.
Four numbers describe every request: the authorized maximum, the cleared prices, the metered actuals, and the refund. Here is a worked example, using the same shape of numbers you will see on a real receipt.
The worked example
- You send a request with a 4,000-token budget. In this equal-service example the winner asks $0.42 per million input-token equivalent and an independent rival sets a $0.55 clear. The receipt records
price_setter: runner-up. The 402 response quotes your worst case: the full budget at cleared rates, say $0.0022. - You authorize $0.0022 of USDC. That is a ceiling, not a bill.
- The reply actually uses 1,000 input and 500 output tokens. The meter charges $0.0009 at the cleared rates.
- The difference, $0.0013, refunds automatically. You never claim it; it just comes back.
Where the charged money goes
The router's fee is 7% of cleared volume, taken out of the provider's side at payout. There is no spread-taking: the router never buys low from providers and sells high to you, so its only way to earn more is more cleared volume. Providers receive one netted payout per hourly epoch.
The price-setter, itemized
Your receipt names whether an independent rival, a disclosed phantom reserve, the winner's own bid, or a session lock set the tariff. Analytics reconstructs the auction, shows rival counterfactuals, and itemizes any premium over the winner's ask plus the fee split. The worked example is a rival-priced case; not every second-score/3 clear has that shape.
- router/src/api/metering.ts token counting on the live stream
- router/src/operations/analytics.ts per-request auction counterfactuals
- router/src/money/payouts.ts epoch netting and the fee