The gateway

Burn vouchers to run inference on a cluster's model. Burns are final.

Burning vouchers

The gateway is the only place vouchers become worth something concrete. burn_for_inference burns vouchers and the gateway executes a model API call against the cluster's model. Burns are final: there is no refund, no reversal and no way to unburn a voucher you spent on a prompt that returned garbage.

Cost per call

Planned voucher cost per call:

cost = max(0.25, price_out * 0.012) vouchers per call

price_out  the model's live output price, USD per 1M tokens
0.25       the minimum charge, in vouchers

Examples

Examples at current catalog prices: Claude Opus 5.5 at $20.00 out costs 0.24, so the 0.25 floor applies. GPT-6 Astra Pro at $50.00 out costs 0.60 vouchers per call. Each voucher covers 1,000 output tokens at the model's live API price, so a 0.60 voucher burn buys about 1,000 output tokens on Astra Pro. If the API price feed moves, the same voucher buys more or fewer tokens. If the gateway is down, vouchers buy nothing at all.