How MPPI works
Inference is measured in tokens. Payment settles in USDC. Follow the work from a funded session to confirmed collection.
On this page
Exchange inference tokens. Settle in USDC.
Choose a provider and accept its token prices. Fund a session in USDC, then let the agent SDK handle signed payment messages alongside your inference requests.
Inference tokens measure a model's input and generated output. MPPI prices that usage in USDC. Each request keeps the prices it accepted, even if the agent provider later changes its catalogue.
From inference to settlement
01Discover
- Inference
- Choose a provider and model.
- Payment
- Read identity and accepted prices. Request unpaid setup on Control.
- On Arc
- Verify the provider and trusted deployment. No inference deposit yet.
The map connects stages of inference to payment events. Collection can happen while a session stays open; each token does not need its own blockchain transaction. The map does not represent measured timings.
Before work: the provider measures input requirements and an output budget. MPPI checks deposited funds and signed payment authorization, including funds reserved for other requests in the same session. Insufficient funds stop admission; a later top-up does not restart that request.
During work: the provider reports cumulative usage at least every 1,000 output tokens, earlier when needed, and at completion. The SDK renews authorization as work progresses. If coverage runs out, the provider must stop within the covered budget. Unresolved work keeps its reservation.