GKE offers two cluster modes: Autopilot and Standard. The mode you choose determines who manages the nodes, how you’re billed, and how much control you have over the underlying infrastructure.

Mode Comparison

AspectAutopilotStandard
Node managementGoogle manages all nodesYou manage node pools
BillingPod resource requests + $0.10/hr cluster management feeCompute Engine VMs + $0.10/hr cluster management fee
Node visibilityNodes exist but are abstractedFull access to node configuration
Idle node behaviorCan scale to zero nodes when no workloads are runningYou configure node pool minimums
Machine type selectionGoogle optimizes per workloadYou choose machine types
OS imageGoogle-managed (Container-Optimized OS)You choose: COS, Ubuntu, etc.
SSH to nodesNot availableAvailable
Node-level SSH/customizationNot supportedFull access
Security postureHardened by defaultYou configure hardening
Kubernetes versionAuto-managed within release channelAuto or manual
Best forMost workloads, hands-off operationsCustom hardware, specific OS, daemon sets

Decision Flow

flowchart TD
    START["Need a GKE Cluster"] --> Q1{"Do you need custom node OS,<br>SSH access, or unsupported hardware?"}
    Q1 -->|Yes| Q2{"Do you need daemon sets<br>on every node?"}
    Q1 -->|No| AUTO["Choose Autopilot"]
    Q2 -->|Yes| STD["Choose Standard"]
    Q2 -->|No| Q3{"Do you need to control<br>exact node count and sizing?"}
    Q3 -->|Yes| STD
    Q3 -->|No| AUTO

    style AUTO fill:#4CAF50,color:#fff
    style STD fill:#2196F3,color:#fff

When to Choose Autopilot

Autopilot is the right choice for most workloads. Choose it when:

  • You want to focus on application code, not infrastructure management
  • Your workloads use standard container images
  • You want predictable, per-pod billing
  • You don’t need SSH access to nodes
  • You want security hardening by default
  • You want automatic scaling without configuring node pools

Key Insight: In Autopilot, you define what your pods need (CPU, memory) and Google provisions the right nodes. You never think about VMs.

When to Choose Standard

Choose Standard when you need:

  • Hardware or accelerator configurations unavailable or constrained in Autopilot
  • Custom OS images or kernel modules
  • SSH access to nodes for debugging
  • DaemonSets that need host-level access or unsupported node privileges
  • Exact control over node count and machine types
  • Spot/Preemptible VMs for batch workloads (though Autopilot now supports Spot pods)
  • Privileged containers or host-level access

Cost Comparison

Standard Cluster Costs

Total = Cluster Management Fee + (Node VM Costs × Node Count)

As of May 2026:

ComponentCost
Cluster management fee73/month)
GKE free tier$74.40/month credit per billing account for Autopilot or zonal Standard cluster fees
Regional Standard clustersCluster fee applies; free tier credit does not apply
Node VMsCompute Engine pricing (varies by machine type)

Example Standard cluster cost (3 × e2-medium in us-central1):

ItemMonthly Cost
Cluster fee$73.00
3 × e2-medium ($24.27/mo each)$72.81
Total~$146/month

Autopilot Cluster Costs

Total = Cluster Management Fee + Sum of running pod CPU, memory, and ephemeral storage requests

As of May 2026, for default general-purpose Autopilot workloads in us-central1:

ResourceOn-Demand PriceSpot Price
CPU$0.0445/vCPU-hour$0.0133/vCPU-hour
Memory$0.0049225/GiB-hour$0.0014767/GiB-hour
Ephemeral storage$0.0001389/GiB-hour$0.000076395/GiB-hour
Cluster fee$0.10/hour, offset by the monthly GKE free tier credit when eligibleSame

Note: Autopilot pricing is based on what you request for your pods, not actual usage. Setting accurate resource requests directly affects your bill.

Example Autopilot cost (5 pods, each 0.5 vCPU, 1 GB memory):

ItemMonthly Cost
5 × 0.5 vCPU × $0.0445/hr × 730hrs~$81.21
5 × 1 GiB × $0.0049225/hr × 730hrs~$17.97
Cluster fee~$73/month before applicable GKE free tier credit
Total~172/month without it

Cost Decision Guide

ScenarioCheaper Option
Small clusters (1-5 nodes) with variable workloadsAutopilot (pay for pod requests, no idle node cost)
Large clusters with consistent, high utilizationStandard (pack pods tightly on VMs)
Development/staging (off-hours)Autopilot (can scale to zero nodes, no idle node cost)
GPU workloadsDepends on GPU type availability
Batch/spot workloadsBoth offer Spot options — compare

Zonal vs Regional Clusters

Autopilot clusters are regional. For Standard clusters, you choose whether the control plane is zonal, multi-zonal, or regional.

AspectZonalRegional
Control planeRuns in one zoneReplicated across 3 zones
NodesIn one zoneDistributed across 3 zones
AvailabilitySingle zone failure = cluster downSurvives single zone failure
CostLower node cost for small clustersHigher VM cost if you run nodes in multiple zones; cluster management fee is the same
Recommended forDevelopment, testingProduction workloads
flowchart LR
    subgraph Zonal["Zonal Cluster"]
        Z1["Zone: us-central1-a"]
        CP1["Control Plane"]
        N1["Nodes"]
    end

    subgraph Regional["Regional Cluster"]
        Z2a["Zone: us-central1-a"]
        Z2b["Zone: us-central1-b"]
        Z2c["Zone: us-central1-c"]
        CP2["Control Plane (replicated)"]
        N2a["Nodes"]
        N2b["Nodes"]
        N2c["Nodes"]
    end

Tip: Use regional Standard clusters for production when you need control plane high availability. Size node pools intentionally because regional clusters can create nodes across multiple zones.

Multi-Zonal Clusters

There is also a multi-zonal option where the control plane runs in one zone but nodes span multiple zones:

FeatureZonalMulti-ZonalRegional
Control plane replicas1 zone1 zone3 zones
Node distribution1 zoneMultiple zones3 zones
Control plane HANoNoYes
Node HANoYesYes

Converting Between Modes

You cannot convert an existing Standard cluster to Autopilot or vice versa. To switch modes:

  1. Create the new cluster in the desired mode
  2. Migrate workloads using Kubernetes manifests or Helm charts
  3. Update DNS/load balancer to point to the new cluster
  4. Delete the old cluster

Warning: Plan the migration carefully. Test the new cluster with production-equivalent workloads before cutting over.

TL;DR

  • Autopilot: Google manages nodes, you pay per pod, security hardened by default — best for most workloads
  • Standard: You manage nodes, you pay per VM, full control — best for custom hardware or node-level requirements
  • Autopilot clusters are regional; use regional Standard clusters for production when you need Standard mode
  • You cannot convert between modes — choose wisely upfront
  • Cost comparison depends on utilization: Autopilot favors variable workloads, Standard favors consistently packed nodes

Resources