Concepts
How compute is billed
Per-second billing from the node agent’s own record of what ran, what a pod costs running and stopped, and how storage is metered.
Per second, from evidence
Pods, serverless workers and dedicated endpoint replicas are billed per second. The seconds come from the agent on the machine that ran the container: it keeps a journal of when each container started and stopped and reports it to the platform. Only whole seconds the journal proves are billed. Time the platform cannot prove (a machine that stopped reporting, for example) is not billed.
Prices are set per second and shown per hour. A price of $0.054 per hour is $0.000015 per second, which is exactly 15,000 nano-USD.
Running, stopped and terminated pods
| State | Compute | Container disk | Pod volume | Network volumes |
|---|---|---|---|---|
| Running | billed | billed | billed | billed |
| Stopped | not billed | not billed | billed | billed |
| Terminated | not billed | gone | deleted unless kept | billed until deleted |
GET /v1/pods/estimate returns a configuration's price while running and while stopped, line by line, from the price book in force.
Storage
Storage is metered in GB-seconds: a volume's size times the seconds it exists. A network volume is billed from creation until you delete it, whether or not a pod uses it.
Serverless workers and dedicated endpoints
A serverless worker is billed per second while its container runs, including the time it waits for jobs before its idle timeout removes it; an endpoint with no workers costs nothing. A dedicated endpoint replica is billed per second from the moment it is ready to serve, not while it loads; requests to a dedicated endpoint are not charged per token.
Interruptible pods
A pod created with pricing_class: interruptible costs less and may be stopped (preempted) when on-demand work needs the capacity; it then waits in the queue and runs again when there is room. On-demand pods are never preempted.