GKE pricing is a cluster management fee per cluster-hour plus, in Standard mode, the Compute Engine price of every node, and in Autopilot mode a price per vCPU-hour, GiB-hour and GiB of ephemeral storage that the pods request, with a surcharge for performance and accelerator compute classes; around the cluster sit the load balancers, the public addresses, the egress, Backup for GKE per namespace and GiB, and the extended support fee for a minor version kept past its standard period. The free tier credits one zonal or Autopilot cluster's management fee per billing account. On the bill, the nodes dominate in Standard mode and the requested pods in Autopilot, and the cluster fee is the line that stays whether the cluster has one node or a hundred.
Pricing dimensions
Google sells a GKE cluster twice over: the control plane as a flat fee, and the capacity either as nodes you size or as pod requests Google sizes for you. Every item has its own SKU, and a node pool in Standard mode is Compute Engine with GKE labels.
| Dimension | Unit | What drives it | Optional |
|---|---|---|---|
| Cluster management fee | USD 0.10 per cluster-hour, by the second, zonal, regional or Autopilot alike; the free tier credits USD 74.40 a month per billing account, one zonal or Autopilot cluster | Every cluster, with or without nodes | No |
| Standard mode nodes | Compute Engine vCPU-hours and GiB-hours per machine family, disks, Local SSD and GPUs at their rates; sustained use, Spot and committed use discounts as on Compute Engine | The node pools and their sizes, whether pods fill them or not | No in Standard mode |
| Autopilot general-purpose | vCPU-hour, GiB-hour and GiB-hour of ephemeral storage per pod request, rounded to the resource ratios Autopilot allows | Every pod's requests, not the node | No in Autopilot |
| Autopilot compute classes | Balanced and Scale-Out at their own rates per vCPU and GiB; Performance and Accelerator classes at the Compute Engine price of the node plus a premium per vCPU, GiB and GPU | Pods that select a class | Yes |
| Autopilot discounts | Spot pods at about 70 percent off; committed use discounts over one or three years on Autopilot vCPU and memory, or the flexible CUD across compute | Interruptible and steady pod load | Yes |
| Enterprise edition | Per vCPU-hour of the fleet for the Enterprise tier with fleet management, policy and service mesh | Clusters registered to an Enterprise fleet | Yes |
| Backup for GKE | Per namespace per backup plan per hour, plus GiB-month of backup storage | Protected namespaces and backup sizes | Yes |
| Extended support | An additional cluster management fee per cluster-hour while the minor version sits in the extended support period of the Extended release channel | Clusters kept on an old minor version | Yes |
| Multi-cluster gateway and ingress | Per cluster-hour for the multi-cluster objects plus the load balancers at their rates | Multi-cluster traffic | Yes |
| Around the cluster | Load balancers, forwarding rules, public addresses, NAT and egress at network rates; Container-native monitoring and logging at their rates | Services of type LoadBalancer, ingress, logs | Yes (network pricing, Cloud Logging and Monitoring) |
| Purchase option | Compute Engine commitments on Standard nodes; Autopilot commitments or the flexible CUD on pods | Steady baseline | Yes (Google Cloud committed use discounts) |
The mode is the pricing decision. In Standard mode the cluster pays its nodes as Compute Engine instances, every hour, whether the pods fill them or not; what the cluster costs is the node pools' size, and the pods' requests decide only how many nodes the autoscaler keeps. In Autopilot the cluster pays the pods' requests at a rate per vCPU and GiB that is about twice the E2 rate on Compute Engine, and nothing for the node, the system pods or the unused capacity; what the cluster costs is the sum of the requests, so a pod that requests four times what it uses pays four times. The cluster fee is the same in both modes and in every topology, zonal or regional, and the free tier pays it for one cluster per billing account, which makes the second cluster the first that costs. The extended support fee is the price of staying on a minor version past its standard support, charged as a second management fee per cluster-hour. The mechanics across all three providers are in containers and Kubernetes; how the three managed services' fees compare, in EKS, AKS and GKE fees compared; what a pod costs on a shared node, in what a pod costs.
How it appears in your billing data
In the billing export the cluster fee carries service.description Kubernetes Engine and sku.description Regional Kubernetes Clusters or Zonal Kubernetes Clusters per hour, with Autopilot Kubernetes Clusters for Autopilot; the free tier appears as credits.type FREE_TIER beside it. Standard nodes do not appear under GKE at all: they are Compute Engine rows, E2 Instance Core running in Frankfurt and the rest, with system_labels and the labels goog-gke-node and goog-k8s-cluster-name that GKE sets on every node, which is how a cluster's nodes are told apart from other instances; the node's disks, Local SSDs and GPUs sit under Compute Engine too. Autopilot pods appear under Kubernetes Engine as Autopilot Pod vCPU Requests (europe-west3), Autopilot Pod Memory Requests (europe-west3) and Autopilot Pod Ephemeral Storage Requests (europe-west3), with the class in the name for Balanced, Scale-Out, Performance and Accelerator pods and Spot for Spot pods; usage.unit is hour, gibibyte-hour and gibibyte-hour. Backup for GKE appears as Backup for GKE Management Fee per namespace-hour and Backup for GKE Storage per GiB-month, extended support as Extended Support Kubernetes Clusters per hour.
What the export does not deliver is the cost per namespace, workload or team: Standard nodes are instances, Autopilot pods are sums per cluster and class, and the labels on a pod never reach the export. GKE cost allocation, switched on per cluster, writes the namespace and the pod labels into the export's labels as k8s-namespace and k8s-label/<key> on the Compute Engine rows of Standard nodes and the Autopilot rows, by the pods' requests; without it, the attribution needs a cost tool on the cluster's metrics, as Kubernetes cost data and Kubernetes cost allocation describe. The export's structure is described in the Google Cloud billing export; in the FOCUS export, ServiceName carries Kubernetes Engine or Compute Engine, SkuId the SKU, and the GKE labels sit in Tags (FOCUS columns).
Levers
Configuration: in Standard mode the first lever is the node pool, its machine family and its autoscaling bounds, because the nodes pay whether pods fill them or not; the second is the pods' requests against their usage, because requests decide how many nodes the autoscaler keeps (cluster autoscaler and Karpenter, pod autoscaling and the bill); in Autopilot the requests are the bill itself, and a request set to the usage halves the line where the habit was to double it. The third is the number of clusters, because every cluster beyond the free one pays USD 73 a month of fee before the first pod, and namespaces separate teams cheaper than clusters (non-production clusters).
Purchase option: Compute Engine commitments on the node pools that run around the clock, Spot node pools or Spot pods for everything interruptible (spot and preemptible nodes), Autopilot commitments or the flexible CUD for steady pod load; the cluster fee has no discount but the free tier.
Architecture: Autopilot for clusters whose pods are sized honestly and whose utilisation on Standard would be low, Standard for clusters that pack nodes above 70 percent or need node-level control; regional clusters only where the control plane's availability matters, because the fee is the same but the nodes are three zones; and the version cadence that keeps clusters off the extended support fee. GKE cost allocation on from day one, because the export carries no namespace without it.
Where it goes wrong
- The second cluster costs the fee. The free tier covers one zonal or Autopilot cluster per billing account; every further cluster pays USD 0.10 per hour. Ten development clusters pay USD 730 a month before a pod runs.
- Standard nodes at 30 percent. The pods fill a third of the nodes and the nodes pay in full. Idle and unallocated cluster cost names the measure; bin packing and smaller nodes are the fix.
- Autopilot pays the request. A pod that requests 2 vCPU and uses 0.2 pays 2. The request, not the limit, is the bill; vertical pod autoscaling in recommendation mode shows the gap.
- The cluster stayed on the old version. The Extended release channel adds a second management fee per cluster-hour once the minor version enters extended support. The upgrade ends it.
- A load balancer per service. Every service of type LoadBalancer pays a forwarding rule and an address per hour; an ingress or gateway bundles them (services, load balancers and egress).
- Logs from every pod. Container logs flow into Cloud Logging per GiB by default; a noisy deployment pays more in logs than in compute. Exclusion filters and log levels are the lever.
- No namespace in the export. GKE cost allocation was never switched on, and the month's cost cannot be split by team. It applies from the day it is enabled, not backwards.
Example
List prices in USD for Frankfurt (europe-west3), read on 2026-10-21 from the GKE and Compute Engine pricing pages: the cluster management fee costs USD 0.10 per cluster-hour, USD 73.00 a month, with USD 74.40 of free tier per billing account; an Autopilot general-purpose pod USD 0.0573 per vCPU-hour and USD 0.0063421 per GiB-hour of memory, USD 0.04584 and USD 0.0050737 with a one-year GKE commitment, USD 0.031515 and USD 0.0034882 with three years, as Spot pods USD 0.0172 and USD 0.0019026; the Balanced class USD 0.0831 per vCPU-hour and USD 0.0091933 per GiB-hour; a Standard node e2-standard-4 (4 vCPU, 16 GiB) USD 0.17267112 per hour on Compute Engine, an n2-standard-4 USD 0.250248; Backup for GKE USD 0.014178082 per namespace-hour and USD 0.00007089 per GiB-hour of backup storage.
A Standard cluster with five e2-standard-4 nodes costs USD 630.25 a month of nodes plus USD 73.00 of fee, USD 703.25, and carries 20 vCPUs and 80 GiB; if its pods request 8 vCPUs and 32 GiB in all, the same workload on Autopilot costs 730 × (8 × 0.0573 + 32 × 0.0063421) = USD 482.78 plus the fee, USD 555.78, and on Autopilot Spot pods USD 145.00 plus the fee. The same pods with requests set to their usage of 4 vCPUs and 16 GiB cost USD 241.39 plus the fee on Autopilot, and the Standard cluster could drop to three nodes, USD 378.15 plus the fee. An organisation with 12 clusters pays 11 × 73.00 = USD 803.00 a month of management fees after the free tier, and the same 12 teams in namespaces of two clusters USD 73.00. Backup for GKE on 10 namespaces with 500 GiB of backups costs USD 103.50 plus USD 25.88 a month. The figures show that Autopilot wins when requests are honest and Standard when nodes are packed, that the fee is the line the number of clusters sets, and that Spot pods cut the pod line by 70 percent for load that tolerates it.
Sources
Read on 2026-10-21. Prices: Google Kubernetes Engine pricing, general-purpose machine family pricing. Service: Autopilot pricing model, GKE cost allocation, extended support. Billing data: structure of the standard Cloud Billing data export.