The cluster fee and the free tier credit
Google charges $0.10 per cluster per hour, in one-second increments, whatever the mode, size or topology. The GKE free tier gives $74.40 of credit per month per billing account, the equivalent of one Autopilot or zonal Standard cluster. The credit covers neither regional clusters nor compute, and unused credit does not roll over. The Extended release channel adds a fee during extended support.
Onicost applies the public $0.10 fee to each GKE cluster and does not deduct this credit. Your first zonal or Autopilot cluster can therefore show a control plane cost that Google offsets on the bill. Keep that gap in mind when you compare the two amounts. Google backs a 99.95% SLA on the control plane of Autopilot and regional Standard clusters, and 99.5% on zonal Standard clusters. A zonal cluster covered by the credit therefore also changes your availability target.
Autopilot: you pay for your requests
In the Pod-based model, the Autopilot default, Google bills by the second the CPU, memory and ephemeral storage your pods request, with no minimum duration. In Iowa (us-central1), the page lists $0.0445 per vCPU-hour and $0.0049225 per GiB-hour. Spot pods sit at $0.0133 and $0.0014767. The Balanced class rises to $0.0645 and $0.0071354. Prices change by region, and Onicost reads yours.
Google also rewrites some requests. On a non-bursting cluster, CPU rounds up to the next quarter of a vCPU. The general-purpose class keeps 1 to 6.5 GiB per vCPU, Balanced 1 to 8 and Scale-Out exactly 4. A pod without requests gets 0.5 vCPU and 2 GiB. In Iowa, that makes about $23.43 per 730-hour month for CPU and memory, against about $9.92 for 0.25 vCPU and 0.5 GiB. We computed both amounts from list prices.
Failing pods cost money too. Google bills scheduled pods, including those stuck in CrashLoopBackOff or ImagePullBackOff. It does not bill unmodified system DaemonSets, unallocated space or pods waiting to be scheduled.
How Onicost reads Autopilot
Onicost recognises Autopilot nodes, their class and their Spot flag. It rebuilds the billed size of each pod from its requests with the rules above, then applies the price of the class in your region. In the Serverless table, the gap between that size and real usage appears under Billed but not used. When all the pods of a workload sit one step or more above their recommended size, an Oversized serverless pods recommendation opens.
Two limits apply. Bursting cannot be seen from the cluster, so Onicost uses the non-bursting rule, which can overestimate a small pod by 0.25 vCPU at most. Pods that ask for specific hardware, through Performance, Accelerator or GPUs, are billed by the node with a management premium. Onicost prices those nodes as machines and leaves the premium aside.
Standard: pools, Spot and free capacity
In Standard mode, you pay each node at its Compute Engine price until you delete it, per second with a one-minute minimum. Empty capacity stays on the bill. Onicost shows it on the Unused line and groups nodes by GKE node pool. It simulates pod placement to suggest a smaller pool or another machine type, as estimates.
Spot nodes carry the label cloud.google.com/gke-spot=true, and Onicost reads it to apply the Spot price. Google states that they have no expiration time, that they are preempted when Compute Engine needs the resources, and that they get 30 seconds of graceful termination by default. Google recommends them for stateless, batch or fault-tolerant workloads. Onicost proposes Spot to the Deployments it finds eligible and recalls that it knows no interruption rate.
Committed use discounts
For Autopilot in Iowa, the page lists $0.0356 per vCPU-hour with a one-year Kubernetes Engine commitment and $0.024475 over three years. Compute Flexible commitments list $0.03204 and $0.02403. Onicost computes its Kubernetes costs at public prices, less the global discount you enter. The savings of node recommendations ignore your commitments too. The Cloud space lists your commitments and their coverage once the Google Cloud account is connected.
Start with a cluster whose requests nobody has reviewed
Install the agent with the Helm command Onicost gives you, then open the Serverless table, sorted from the most expensive workload down. The Free plan handles one cluster up to 32 cores. On Autopilot, each pod counts there for its billed vCPUs. Team starts at €39 per month before tax with detailed recommendations and the scheduler, and you can try it for 14 days without a card. Amounts stay estimates without network traffic, and Onicost guarantees no saving.
Frequently asked questions
Does Onicost support GKE Autopilot?
Yes. Pods billed by the pod are priced at their billed size, at the price of their class and region. The Performance and Accelerator classes stay priced as machines.
Does the $74.40 credit appear in Onicost?
No. Onicost shows the public fee of $0.10 per hour for each cluster. The credit applies on your Google Cloud bill, per billing account.
Is Autopilot cheaper than Standard?
That depends on how full your nodes are. Autopilot bills requests, and Standard bills nodes even when half empty. Compare the Unused line of your Standard clusters with the part billed but not used on Autopilot.