The idle bill is whatever was provisioned ahead of the traffic. Charges metered per token or per call stop when the traffic stops. Capacity bought in advance is charged for every hour it is held, whether or not a request arrives.
Akka scales elastically with scale-to-zero, so idle capacity costs nothing. Orchestration, agents, memory, streaming, endpoints and guardrails draw on one shared pool, so a single pool covers the peak of all of them.
A service billed per unit of work charges nothing between requests. Model inference is priced per token and an API gateway is priced per call, so an agent receiving no traffic generates no charge on either.
Provisioned capacity is different in kind. A throughput unit on an event stream, a node in a vector index, a minimum instance count on a hosted endpoint, and a memory store sized for peak all bill for the hours they exist. The charge follows the reservation, and the reservation is made before the traffic is known.
Agent traffic is uneven. Overnight document processing, a support agent following business hours, and a seasonal claims spike all leave long stretches with no requests arriving. Capacity provisioned for the peak is charged through every one of those stretches.
Akka scales elastically to 10 million agentic transactions per second and back to zero. The capacity is shared, so the peak of the streaming work and the peak of the agent work draw on one pool sized once.
A paused run costs storage. A run waiting on a human decision holds its state in the journal and releases the compute it was using, so thousands of waiting runs consume no capacity while they wait.
Scaling up from zero adds start-up time to the first request after an idle period. The trade is between that latency and the cost of holding capacity through the idle window, and the setting is chosen per service.
Akka bills on platform tier and service cores. In-VPC tiers disclose platform cores, which are Akka infrastructure overhead, separately from service cores, which carry your workloads. Akka charges for service cores.
State persists in the event journal while the compute is released. An agent that has not been called for a week returns holding everything it accumulated.
The Sandbox tier installs a single region in your own VPC with platform-core overhead minimized. Sandbox regions carry no HA, no DR and no SLA, and they are not intended for production workloads.
