All Products
Search
Document Center

Agent Sandbox:Billing Overview

Last Updated:Sep 18, 2026

Agent Sandbox charges for the computing resources that it provides and for the resources of dependent Alibaba Cloud services. Review these billing rules to estimate and control your costs.

Billing components

The total fees for Agent Sandbox consist of the Agent Sandbox compute fee and the cloud service resource fee.

  • Agent Sandbox compute fee: The fee for the vCPU, memory, and disk computing resources provided by Agent Sandbox.

  • Cloud service resource fee: The fee for the resources of other Alibaba Cloud services that you use. The corresponding service charges this fee.

Billing method

Agent Sandbox supports the pay-as-you-go billing method. The following rules apply:

  • Billing granularity: Resources are billed by the second. A usage period of less than 1 second is counted as 1 second.

  • Billing cycle: A bill is generated every hour.

Compute classes (QoS)

Agent Sandbox provides three Quality of Service (QoS) compute classes. The three classes share the same billing rules: per-second billing, a minimum billable period of 1 second, and hourly billing. Only the unit prices differ. Select a class based on your workload's requirements for predictable compute performance.

QoS class

Edition support

Compute characteristics

Typical scenarios

Economy

Flash Edition and Container Edition

Cost-effective

Allows compute performance to fluctuate and has lower unit prices than the other classes.

Agent development and debugging, Tool Use workflow validation, and infrequent Code Interpreter workloads.

Default

Recommended

Flash Edition and Container Edition

Balanced performance and scalable throughput

Provides relatively stable performance, supports very large-scale deployments, and delivers low-latency responses for enterprise production environments.

Code Interpreter production services, highly concurrent agent execution for consumer-facing applications, large-scale sampling and reinforcement learning (RL) simulation, and multi-tenant sandbox isolation.

Performance

Supported only by the Flash Edition

Guaranteed resource delivery

A dedicated resource pool reduces compute performance fluctuations and supports wake-up from light hibernation within 1 ms.

Production-grade RL training, quantitative finance and real-time decision-making, and long-running multi-agent workflows.

Pricing details

The prices in the following tables are reference prices for cost estimation. Actual charges are based on your bill in the Expenses and Costs console. Available regions, resource specifications, and unit prices may change; updates are reflected in this topic.

International site pricing

Region

QoS class

vCPU price (USD/vCPU/second)

Memory price (USD/GiB/second)

Disk price (USD/GiB/second)

Chinese mainland

Economy

USD 0.0000026/second

(USD 0.00936/hour)

USD 0.00000128/second

(USD 0.004608/hour)

USD 0.0000000886/GiB/second

(USD 0.00031896/GiB/hour)

Default (recommended)

USD 0.0000034/second

(USD 0.01224/hour)

USD 0.00000167/second

(USD 0.006012/hour)

Performance

USD 0.0000052/second

(USD 0.01872/hour)

USD 0.0000026/second

(USD 0.00936/hour)

China (Hong Kong) and regions outside China

Economy

USD 0.00000381/second

(USD 0.013716/hour)

USD 0.00000191/second

(USD 0.006876/hour)

USD 0.0000000703/GiB/second

(USD 0.00025308/GiB/hour)

Default (recommended)

USD 0.00000496/second

(USD 0.017856/hour)

USD 0.00000248/second

(USD 0.008928/hour)

Performance

USD 0.00000759/second

(USD 0.027324/hour)

USD 0.00000386/second

(USD 0.013896/hour)

Billing rules

Resource specifications

Fees are charged based on the vCPU and memory specifications that you set when you create a sandbox instance. The Flash Edition strictly validates the resource specifications. The Container Edition supports flexible specification configuration in Kubernetes YAML files. For the Container Edition, if a vCPU or memory specification is not supported, the cluster automatically normalizes it and bills based on the normalized specification.

Sandbox instance states

A sandbox instance supports the following states, and the billable items vary based on the state:

State

Edition support

State description

Billable items

Active

Flash Edition and Container Edition

The sandbox instance runs as expected and handles user requests and tasks in real time.

vCPU, memory, and disk

Deep hibernation

Flash Edition and Container Edition

The entire sandbox instance state, including memory, is saved as a snapshot in persistent storage. The instance consumes only storage resources and no vCPU or memory resources. After the instance is resumed, it continues to run from where it stopped.

Disk

Light hibernation

Supported only by the Flash Edition

An idle sandbox instance automatically enters the light hibernation state through request sensing and memory snapshot technologies to reduce costs. An instance is considered idle if one of the following conditions is met:

  • The CPU usage is less than 0.01 core.

  • The network traffic is less than 1,000 Bytes/s.

  • The instance receives no requests.

Memory and disk

Snapshot saved

Flash Edition: Snapshot

Container Edition: Checkpoint

The state of a running sandbox instance is captured, and a snapshot of the current memory and file system is persisted to storage space. New instances can be cloned from the snapshot to quickly replicate and distribute environments.

Disk

Destroyed

Flash Edition and Container Edition

The sandbox instance is permanently terminated, and all vCPU, memory, and storage resources are released.

None

Active state

When a sandbox instance is in the active state, you are charged for the vCPU, memory, and disk (rootfs) resources. The rootfs is the temporary storage space of the container root file system. A free disk allowance of 15 GiB applies.

Fee for a single sandbox = (Number of vCPUs × vCPU unit price + Memory size × Memory unit price + Disk usage × Disk unit price) × Billing duration
Disk usage = max(0, Disk size - 15) GiB

Deep hibernation state

When a sandbox instance is in the deep hibernation state, you are not charged for the vCPU and memory resources. You are charged only for the disk based on the amount of storage space occupied during deep hibernation. The free disk allowance of 15 GiB does not apply during deep hibernation.

Free disk allowance: none
Deep hibernation storage = Instance memory size + Instance disk size
Fee for a single sandbox = Deep hibernation storage × Disk unit price × Hibernation duration

Light hibernation state

When a sandbox instance is in the light hibernation state, you are not charged for the vCPU resources. You are charged only for the memory and disk (rootfs temporary storage space) resources. A free disk allowance of 15 GiB applies. Light hibernation is supported only by the Flash Edition.

Fee for a single sandbox = (Memory size × Memory unit price + Disk usage × Disk unit price) × Billing duration
Disk usage = max(0, Disk size - 15) GiB

Snapshot saved state

When a sandbox instance is in the snapshot saved state, you are not charged for the vCPU and memory resources. You are charged only for the disk based on the amount of storage space occupied by the snapshot. The free disk allowance of 15 GiB does not apply to snapshots.

Free disk allowance: none
Snapshot storage = Instance memory size + Instance disk size
Fee for a single sandbox = Snapshot storage × Disk unit price × Snapshot retention duration

Destroyed state

A destroyed sandbox instance no longer incurs Agent Sandbox instance fees.

Cloud service resource fees

If you use the resources of other Alibaba Cloud services alongside Agent Sandbox, you are charged for these resources based on the billing rules of the corresponding services. The corresponding services collect the fees.

Overdue payments and resource release

  • Pay-as-you-go bills are typically generated every hour. When an insufficient account balance causes an overdue payment, the system sends an overdue payment notification.

  • If your account has an overdue payment, sandbox instances may be automatically stopped. Stopped instances cannot run workloads.

After your tasks are complete, destroy unused sandbox instances. Set a proper timeout period for production tasks to prevent unexpected fees caused by long periods of idleness.