EKS pricing is a fee per cluster hour that rises sixfold with extended support for old Kubernetes versions, plus the EC2 instances or Fargate pods the workloads run on, plus a management fee per instance hour once EKS Auto Mode runs the nodes; ECS has no cluster fee and charges only the compute beneath it, and Fargate charges per vCPU-second and GB-second a task or pod requests. On the bill the nodes almost always dominate, not the cluster; what makes a cluster dear is the instances it does not fill, and the load balancers, NAT gateways and logs around it.
Pricing dimensions
The three services share one logic: the orchestrator is either free (ECS) or a small fixed fee (EKS), and the money goes into the compute, billed either as an instance under EC2 or as requested vCPU and GB under Fargate. Every item carries its own usage type, and the instances sit not under EKS or ECS but under EC2.
| Dimension | Unit | What drives it | Optional |
|---|---|---|---|
| EKS cluster | Hour per cluster, at any size | Every cluster, empty or full, from creation to deletion | No on EKS |
| EKS extended support | Hour per cluster, five times the cluster fee on top | Every cluster on a Kubernetes version past 14 months of standard support, automatically for up to 12 more months | Yes, through an upgrade or by disabling it |
| EKS Auto Mode | Hour per managed EC2 instance, by instance type about an eighth of the instance price | Every instance Auto Mode launches and manages, on top of the instance itself | Yes |
| EKS Provisioned Control Plane | Hour per cluster by tier (XL to 8XL), on top of the cluster fee | Clusters that need reserved control plane capacity | Yes |
| EKS Capabilities | Hour per enabled capability plus hour per managed resource (Argo CD application, ACK resource) | Every capability in the account | Yes |
| Nodes on EC2 | Second per instance type, under EC2 | The node groups and node pools with their volumes | No (EC2 pricing, EBS pricing) |
| Fargate | vCPU-second and GB-second per task or pod, from the image pull to the end, one minute minimum | The requested vCPU and GB in the permitted combinations; Linux x86, Linux ARM and Windows at their own rates, Windows with an OS surcharge per vCPU | Yes, instead of EC2 |
| Fargate ephemeral storage | GB-hour above the 20 GB every task brings | Tasks with configured storage up to 200 GB | Yes |
| Fargate Spot | vCPU-second and GB-second at the variable spot price, ECS tasks on Linux only | Interruptible tasks; interruption with two minutes' notice | Yes |
| ECS Managed Instances | Hour per managed EC2 instance by type, on top of the instance | Instances ECS Managed Instances launches and manages | Yes |
| Around the cluster | Load balancer per hour and LCU, NAT gateway per hour and GB, public IPv4 addresses per hour, Container Insights logs per GB, ECR per GB-month | Every ingress, every egress, every line of log | Yes (load balancing and IPv4, data transfer, CloudWatch) |
| Purchase option | Compute Savings Plan on EC2 nodes and Fargate; Reserved Instances and EC2 Instance Savings Plans on nodes; nothing on the cluster fee | One- or three-year term | Yes (AWS Savings Plans) |
The EKS cluster fee applies per hour whatever the number of nodes and pods, and it is small only when the cluster is large; ten clusters for ten environments pay it ten times. A Kubernetes minor version sits under standard support for 14 months, then enters extended support automatically for up to 12 months, which lifts the cluster fee to six times; whoever does not upgrade is moved to the oldest still supported version at the end of the period. Auto Mode takes over node pools, autoscaling and the add-ons for storage, networking and load balancing and charges per instance hour an amount that grows with the instance type, about twelve percent of the on-demand price in Frankfurt; ECS Managed Instances charge the same model for ECS. Both fees come on top of the instance and replace neither Reserved Instances nor Savings Plans on it.
Fargate replaces the node with the task: you request vCPU and GB in the permitted combinations, from a quarter vCPU with 0.5 GB to 16 vCPU with 120 GB, and pay both per second from the moment the container image is pulled until the task ends, one minute minimum. There is no idle in the sense of a half-empty node, but a task that requests 1 vCPU and uses 0.1 pays 1 vCPU; rightsizing moves from the instance to the task definition. Windows tasks pay a surcharge per vCPU-hour for the licence, ARM tasks a lower rate, and Fargate Spot for interruptible ECS tasks the variable price. ECS itself charges no fee, neither for the cluster nor for tasks on EC2.
How it appears in your billing data
In the Cost and Usage Report the EKS lines carry the product code AmazonEKS and lineItem/ResourceId the cluster's ARN; the cluster fee appears as EUC1-AmazonEKS-Hours:perCluster with the operation CreateOperation, extended support as EUC1-AmazonEKS-Hours:extendedSupport with the operation ExtendedSupport on a line of its own beside the cluster fee, the Provisioned Control Plane as EUC1-AmazonEKS-Hours:XL-ProvisionedTier and Auto Mode as EUC1-EKS-Auto:m7i.large-management-hours per managed instance, the prefix being the region code, EUC1 for Frankfurt. The EKS Capabilities appear as EUC1-AmazonEKSCapabilities-ArgoCD-Hours:perCapability and -CR-Hours:perCustomResource.
Fargate sits under the orchestrator's product code, AmazonECS for tasks and AmazonEKS for pods, as EUC1-Fargate-vCPU-Hours:perCPU and EUC1-Fargate-GB-Hours, with ARM- or Windows- after Fargate- for the other platforms, EUC1-Fargate-Windows-OS-Hours:perCPU for the licence surcharge and EUC1-Fargate-EphemeralStorage-GB-Hours for storage above 20 GB; lineItem/ResourceId names the task or pod. ECS Managed Instances appear as EUC1-ECS-Managed-Instances:m7i.large-management-hours under AmazonECS. The nodes themselves sit under AmazonEC2 as EUC1-BoxUsage:m7i.large with their volumes, and nothing in the report ties them to the cluster except the tag the node group or Karpenter gives them; without eks:cluster-name or a cluster tag of your own, the cluster in the report is only its fee.
What the report therefore does not show per cluster: the cost per pod, namespace or team. For that there is split cost allocation data, which splits the instance cost by the pods' requests and writes it as lines of its own into the report, or a tool inside the cluster; what a pod costs is described in what a pod costs, and how the data is produced in Kubernetes cost data. The report's structure is described in the AWS Cost and Usage Report. In the FOCUS export, ServiceName carries Amazon Elastic Container Service for Kubernetes or Amazon Elastic Container Service and x_ServiceCode AmazonEKS or AmazonECS; cluster fee, Auto Mode and Fargate are ChargeCategory Usage, and Fargate under a Compute Savings Plan PricingCategory Committed (FOCUS columns).
Levers
Configuration: the number of clusters is the first lever on EKS, because each carries its fee and its fixed costs; namespaces per environment instead of clusters per environment is the rule non-production clusters describes. The Kubernetes version is the second: an upgrade before the end of standard support saves five times the cluster fee per hour. Node utilisation is the third and the largest, because the instances under EC2 pay for the cluster; node pools and instance selection and cluster autoscaler and Karpenter are the pages for it. On Fargate the requested vCPU and GB per task definition are the lever, because they, not the usage, are paid.
Purchase option: a Compute Savings Plan covers EC2 nodes and Fargate at once and is therefore the form that fits clusters whose mix changes; Reserved Instances and EC2 Instance Savings Plans cover only nodes of one type or family. No commitment covers the cluster fee, Auto Mode or extended support. Which form fits is the subject of commitment strategy, and the split between baseline and burst of capacity planning and headroom.
Architecture: Fargate for workloads that run rarely or vary strongly, EC2 nodes for steady load, because Fargate costs more per vCPU and GB than the same capacity as an instance and moves rightsizing to the task; spot and preemptible nodes for interruptible load. Auto Mode and Managed Instances buy operations for a surcharge that is weighed against the time a team otherwise spends on node pools and add-ons. The three providers' fees are set side by side in EKS, AKS and GKE fees compared, the dimensions in containers and Kubernetes.
Where it goes wrong
- The cluster fee rises sixfold overnight. With the end of a version's standard support, EKS switches extended support on, automatically. Lines with
AmazonEKS-Hours:extendedSupportname the clusters; the upgrade ends them. - Ten clusters, ten fees. Dev, staging, QA and preview as clusters of their own each pay the fee, the ingress, the NAT gateway and the minimum number of nodes. Count the
AmazonEKS-Hours:perClusterlines. - The cluster is half empty. The nodes sit under EC2 and pay in full whether pods run on them or not. The sum of requests against the allocatable capacity per node shows the idle (idle and unallocated cluster cost).
- Fargate pays the request. A task with 4 vCPU and 8 GB that uses 0.5 vCPU pays 4 vCPU per second. Container Insights shows the usage per task; the task definition is where the correction goes.
- Windows tasks carry a licence surcharge. Fargate on Windows pays the operating system per vCPU-hour on top. Lines with
Fargate-Windows-OS-Hoursshow it. - The nodes cannot be attributed. Instances without a cluster tag appear under EC2 like any other. Node groups and Karpenter NodePools assign tags that are activated as cost allocation tags (cost allocation tags).
- The cluster costs more outside than inside. Load balancers per service, NAT gateway bytes and Container Insights logs sit under other product codes and exceed the cluster fee several times over (services, load balancers and egress).
Example
List prices in USD for Frankfurt (eu-central-1), read on 2026-10-21 from the AWS price list: an EKS cluster costs USD 0.10 per hour, with extended support USD 0.60; EKS Auto Mode costs USD 0.01449 per hour for an m7i.large and USD 0.01174 for an m7g.large, on top of USD 0.12075 and USD 0.0978 for the instance itself; ECS Managed Instances USD 0.014496 and USD 0.011736 per hour for the same types. Fargate costs USD 0.04656 per vCPU-hour and USD 0.00511 per GB-hour on Linux x86, USD 0.03725 and USD 0.00409 on Linux ARM, USD 0.053544 and USD 0.0058765 on Windows plus USD 0.046 per vCPU-hour for the operating system; ephemeral storage above 20 GB USD 0.000132 per GB-hour.
A cluster with three m7i.large nodes costs, in a month of 730 hours, USD 73.00 of cluster fee and USD 264.45 of instances, USD 337.45 in all; on a version in extended support USD 438.00 of cluster fee and USD 702.45 in all; with Auto Mode USD 31.73 of management fee on top, USD 369.18 in all. The same cluster on ECS with EC2 nodes costs USD 264.45, because ECS has no cluster fee. A Fargate task with 1 vCPU and 2 GB costs USD 0.05678 per hour, USD 41.45 a month on Linux x86 and USD 33.16 on ARM; a task with 2 vCPU and 8 GB, the capacity of an m7i.large, costs USD 0.1340 per hour against USD 0.12075 for the instance, about eleven percent more, with no idle and no node. The figures show that the cluster fee is a fifth of the bill on small clusters and more than the nodes under extended support, that Auto Mode raises the instances by about twelve percent, and that Fargate sits slightly above the instance per unit of capacity and is cheaper only where the instance would not be filled.
Sources
Read on 2026-10-21. Prices: Amazon EKS pricing, Amazon ECS pricing, AWS Fargate pricing, AWS Price List, offer AmazonEKS for eu-central-1, offer AmazonECS for eu-central-1. Billing data: Cost and Usage Report columns: lineItem, split cost allocation data, FOCUS 1.0 with AWS columns. Service: Kubernetes versions in Amazon EKS, EKS Auto Mode, AWS Fargate for Amazon ECS.