xpu liveBETA

Agent runtime

Google Cloud

Managed agent runtime with sessions, Memory Bank and observability.

Own model keyPlan dependent
Published plans1
Last checked2026-10-23
Source statusSources reviewed

The essentials

What Google Cloud offers

Google Cloud is listed here as runtime infrastructure for AI agent workloads. Managed agent runtime with sessions, Memory Bank and observability.

Pricing and deployment

Usage-based Agent Runtime: $0.085/vCPU-hour and $0.009/GiB-hour, rounded to the nearest second. Idle time between turns is not billed. Sessions, memory, gateways and model usage are separate.

Choosing this service

Evaluate the execution model, persistence requirements and connected services before choosing a deployment. A platform base fee or compute rate is not a complete model-usage budget.

Specifications below distinguish documented features from information not disclosed in the reviewed sources. Official links and the review date are provided for checking current terms.

Subscription & usage

Choose a plan. Check the scope.

Official pricing

These plans can serve different access methods. Model usage and resource charges may be additional.

Agent Runtime — on-demand metering

Subscription

Usage-based

$0.085/vCPU-hour + $0.009/GiB-hour; storage, operations and models extra.

Included

Managed agent runtime with sessions, Memory Bank and observability.

Extra usage

Agent Compute $0.085/vCPU-hour; Agent Memory $0.009/GiB-hour. Agent storage $0.30/GiB-month plus read/write operations. Other services and model usage are additional.

Full conditions & limits

Pricing

Pricing model
Published
Base subscription amount (not total cost)
0
Currency
USD
Billing interval
Usage only
Subscription scope
$0.085/vCPU-hour + $0.009/GiB-hour; storage, operations and models extra.
Annual commitment / upfront payment
Usage-based Agent Runtime: $0.085/vCPU-hour and $0.009/GiB-hour, rounded to the nearest second. Idle time between turns is not billed. Sessions, memory, gateways and model usage are separate.
Setup fee
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Included usage
Managed agent runtime with sessions, Memory Bank and observability.
Overage rates
Agent Compute $0.085/vCPU-hour; Agent Memory $0.009/GiB-hour. Agent storage $0.30/GiB-month plus read/write operations. Other services and model usage are additional.
Model costs and BYOK conditions
Model API usage and third-party tools are additional unless explicitly included by the selected offer.
Free trial / credits / expiry
Published monthly account allowances: 50 compute hours and 100 GiB-hours of memory; check eligibility and aggregation rules.
Cancellation and refunds
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Tax treatment
Not disclosed in the reviewed official sources; confirm with the provider before purchase.

Resources and limits

CPU allocation and billing basis
Metered CPU; selected regional SKU
RAM
Metered memory; selected regional SKU
GPU / VRAM
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Storage and persistence
Session and Memory Bank charges are separate
Network egress
Network and other Google Cloud services can add charges
Agent limit
Default 100 Agent Runtime resources per project/region.
Concurrency
Default 10 concurrent live bidirectional connections; distinct from ordinary query throughput.
Execution quota
Sandbox code-execution API: 1,000 execute requests/minute/project/region; model and other API quotas are separate.
Maximum session / execution duration
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Request rate limit
Per project/region: 90 Query/StreamQuery per minute, 10 resource writes/minute and 100 session writes/minute.
Continuous execution / idle behavior
Runtime pricing excludes idle time waiting between turns; actual allocation and scaling configuration govern usage.
Retention
Not disclosed in the reviewed official sources; confirm with the provider before purchase.

Plan controls

Support
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Contractual SLA / availability target
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Plan-specific security and compliance
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Eligibility and restrictions
Not disclosed in the reviewed official sources; confirm with the provider before purchase.

Evidence

Official pricing source
Official source
Last checked (UTC date)
2026-10-23
Source interpretation and caveats
Representative plan only, not the complete provider catalog. Prices are a dated reference; check official checkout for current regional terms.
1 fields not disclosed

Price effective date, if explicitly published

Under the hood

Explore the specifications

Open a category for the details. Capabilities may depend on the selected plan.

Company & deploymentService model, use cases and limitations

Company and service

Official website
Official source
Documentation
Official source
Hosting category
Agent runtime
Deployment model
Managed platform; customer deploys and configures the agent workload.
Best suited to
Managed agent runtime with sessions, Memory Bank and observability.
Important limitations
Usage-based Agent Runtime: $0.085/vCPU-hour and $0.009/GiB-hour, rounded to the nearest second. Idle time between turns is not billed. Sessions, memory, gateways and model usage are separate.
Compatibility & toolsModels, memory, MCP and integrations

Compatibility and tools

Frameworks / SDKs
ADK, A2A and documented framework integrations.
Languages / runtimes
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Model compatibility
Model availability and credentials depend on the selected framework and model provider.
Bring your own model API key
Conditional / plan dependent
Memory and sessions
Managed sessions and Memory Bank; retention and charges depend on configuration.
MCP: client / server / hosting distinction
Tool integrations depend on the deployed agent and supported SDK.
Schedules and triggers
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Browser execution
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Code execution / sandbox
Execution facilities depend on the selected runtime and configured tools.
App integrations / authentication
Configure supported tools and provider credentials; third-party services can add charges.
Infrastructure & operationsRegions, scaling, networking and recovery

Infrastructure and operations

Available deployment regions
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Scaling and idle behavior
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Git / deployment workflow
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Private network / static IP / egress controls
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Backups and recovery
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Logs, tracing, metrics and evaluations
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Security & supportIsolation, access controls and support

Security and support

Isolation model
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Secrets management
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
SSO / roles / audit logging
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Compliance evidence and scope
Not disclosed in the reviewed official sources; confirm with the provider before purchase.
Support channels
Not disclosed in the reviewed official sources; confirm with the provider before purchase.

Evidence, not assumptions

Sources & verification

Reviewed 2026-10-23 UTC. “Not disclosed” means a value was not confirmed in these sources, not that a capability is unavailable.