Pricing
You pay per token, in USDC on Arc, at the rate of the model's tier. Input tokens cost half the output rate. The table comes straight from the Arcloak catalog.
| Model | Tier | Input | Output |
|---|---|---|---|
| llama3.2:3b | small | 0.05 | 0.10 |
| qwen2.5:7b | small | 0.05 | 0.10 |
| llama3.1:8b | mid | 0.15 | 0.30 |
| qwen2.5:32b | large | 0.50 | 1.00 |
Each request authorises at most 2 USDC and settles only the tokens delivered, rounded up to the nearest 0.000001 USDC. Nodes receive 90% of every settled amount.
How a request is billed
- Before the request runs, Arcloak estimates the prompt at about four characters per token plus four tokens per message, and prices the prompt plus
max_tokensof output. That price is the ceiling your wallet signs. - If the ceiling would pass the per-request cap,
max_tokensis lowered until it fits. If the prompt alone leaves no room for output, the request is refused withinvalid_requestbefore any payment. - The node streams the answer and reports the tokens it counted. Arcloak bills those counts at the tier rate, rounds up to the nearest 6-decimal unit, never charges more than the ceiling, and never less than one unit.
- Only that amount is settled on chain. The rest of the ceiling never moves.
If the node stops mid-answer, only the delivered part is billed. If nothing was delivered, the authorization is cancelled and nothing is charged.
Live prices
GET https://api.arcloak.net/v1/models returns the same prices per model, with the number of nodes serving each model right now.