GKE Autopilot Pricing: What GKE Autopilot Costs per Pod Request and How It Compares to GKE Standard

GKE Autopilot bills general-purpose pods on what they request, not on the nodes underneath: $0.0445 per vCPU-hour, $0.0049225 per GiB-hour of memory and $0.0001389 per GiB-hour of ephemeral storage in us-central1, charged per second with no minimum duration. Every cluster also pays a $0.10 per hour management fee, and the free tier credits $74.40 a month per billing account, enough for one Autopilot cluster. Because the bill follows requests, oversized requests are the single biggest Autopilot cost, and trimming them saves money the same second.

Every Google rate below was read on the GKE pricing page in September 2026, in USD for the Iowa (us-central1) region. Other US regions differ slightly. The arithmetic is ours, at 730 hours a month, and it is shown so you can redo it for your own fleet.

How is GKE Autopilot billed?

Autopilot uses two billing models, and which one applies depends on what your pods ask for. General-purpose pods, meaning pods on the default Autopilot container-optimized compute platform or on the Balanced and Scale-Out compute classes, use pod-based billing: you pay for the CPU, memory and ephemeral storage in the pod's resource requests, in one-second increments, while the pod is running or being created. Pods that select specific hardware, such as a GPU or a Compute Engine machine series, use node-based billing: you pay for the whole Compute Engine node Autopilot creates, plus an Autopilot management premium.

Google does not charge for pods waiting to be scheduled, for pods that have finished, for unmodified system DaemonSets, for operating system overhead or for unallocated space. It does apply two rules that catch people out: if a pod has no request, Autopilot sets a default, and if a request is below Autopilot's minimums or outside its CPU-to-memory ratios, Autopilot raises it. You are billed for the raised value.

How much does GKE Autopilot cost per vCPU and per GiB?

Two things stand out. The Compute Flexible CUD is cheaper than the Kubernetes Engine CUD on both terms ($0.03204 against $0.0356 per vCPU-hour on one year, $0.02403 against $0.024475 on three), and it also covers Compute Engine, so compare the two before you buy either. And the Balanced class costs about 45 percent more per vCPU than the default platform, so pinning pods to Balanced without a reason is an expensive habit. Google says spot prices can change up to once every 30 days but always sit 60 to 91 percent below the regular price.

What does GKE Autopilot cost for a real workload?

A single pod requesting 0.5 vCPU and 2 GiB of memory costs about $0.0321 an hour, or about $23.43 a month, on demand. Scale that up and the bill looks like this, assuming requests at a 1 vCPU to 4 GiB ratio running all month:

Add $73 a month per cluster for the management fee (minus the $74.40 free credit on one cluster per billing account), plus load balancers, persistent disks and network egress, which are billed separately.

Is GKE Autopilot more expensive than Standard?

Per unit of capacity, yes. Per unit of capacity you actually use, often not. A pod requesting 4 vCPUs and 16 GiB costs about $0.257 an hour on Autopilot. An e2-standard-4 VM with the same 4 vCPUs and 16 GiB costs about $0.134 an hour on demand in us-central1. So a Standard node is about half the price of the same shape on Autopilot, but only if your pods fill it.

The break-even is roughly 52 percent: if your pod requests fill more than about half of each Standard node, Standard is cheaper; below that, Autopilot usually is, because you stop paying for the empty half. Most Standard clusters that were sized once and left alone run well under that line, which is why Google pitches Autopilot as a cost saving. The same logic also works the other way. On Autopilot, the empty space moves inside your pods: a service that requests 2 vCPUs and uses 0.4 is paying for 1.6 vCPUs of nothing, at Autopilot's higher rate.

Does GKE Autopilot have a free tier?

Yes, for the cluster fee only. The GKE free tier gives $74.40 in monthly credits per billing account, which covers the $0.10 per hour management fee of one Autopilot or zonal Standard cluster for a full month. It does not cover pod compute, and unused credit does not roll over. A second Autopilot cluster on the same billing account pays the fee in full.

How do I reduce GKE Autopilot costs?

Start with requests, because on Autopilot they are the bill. Then move the steady part to a commitment and the interruptible part to spot.

1. Rightsize pod requests. Vertical Pod Autoscaler is enabled by default in Autopilot clusters. Create VPA objects in Off mode, read two weeks of recommendations, then switch stable workloads to InPlaceOrRecreate, which resizes running pods without a restart on GKE 1.34.0-gke.2201000 or later. On 500 vCPUs of requests, a 30 percent cut is worth about $7,000 a month on demand. The tools that automate this, and what each costs, are compared in our GKE rightsizing tools guide.
2. Check which compute class your pods use. A pod pinned to Balanced pays $0.0645 per vCPU-hour instead of $0.0445. Keep Balanced for workloads that need it.
3. Commit the baseline. After rightsizing, buy a CUD on what is left. Committing before you rightsize locks in waste for one or three years.
4. Put batch and retryable work on spot pods at $0.0133 per vCPU-hour, with a disruption budget that tolerates preemption.
5. Delete idle clusters. Every cluster beyond the free one pays $73 a month before it runs a pod, and $438 a month once it sits on the Extended channel past standard support, where the fee rises to $0.60 an hour.

Recommendations that nobody owns do not get applied. What works in practice is a weekly list per team with the dollar value next to each workload, and a tool that routes each recommendation to the team that owns the workload instead of dropping it into a shared channel.

Which tools rightsize GKE Autopilot, and what do they cost?

On Autopilot, Google owns the nodes, so only the pod-level half of a rightsizing tool matters. The built-in VPA costs nothing. Of the commercial tools, StormForge Optimize Live lists $0.0041 per requested vCPU-hour pay-as-you-go on its AWS Marketplace card, about $2.99 per vCPU a month, and tunes Java heap, which Google's VPA documentation advises against using VPA for. PerfectScale by DoiT is free up to 300 vCPUs a month. ScaleOps lists a $50,000 a year platform fee plus $9.00 per vCPU. Those are AWS listings; on GKE you buy direct or through Google Cloud Marketplace at a quoted price, so treat them as the anchor for the negotiation. How the two automated pod tools compare is in our ScaleOps vs StormForge breakdown.

Run the numbers against the saving. At 500 vCPUs on demand, StormForge would cost about $1,500 a month against the roughly $7,000 a month a 30 percent request cut is worth. At 100 vCPUs the same cut is worth about $1,400, which is well inside what the free VPA or PerfectScale's free tier can deliver.

How do I see GKE Autopilot cost per team?

Turn on GKE cost allocation for the cluster. It adds the cluster name and namespace to the labels in your Cloud Billing export to BigQuery, and puts node overhead in a separate kube:system-overhead namespace so it does not land in anyone's number. On Autopilot that mapping is unusually clean, because the charge already follows each pod's request. If the same company also runs EKS, the equivalent is covered in our EKS Auto Mode pricing analysis.

What should I check before moving to GKE Autopilot?

1. Node fill on Standard today. Divide total pod requests by total node capacity per node pool. Under about 50 percent, Autopilot is likely cheaper. Over it, Standard with good autoscaling probably is.
2. Requests on every workload. Pods without requests get Autopilot's defaults, and tiny requests get raised to the minimums. Set them before the move, not after the first invoice.
3. Hardware selectors. Anything that pins a GPU or machine series moves to node-based billing plus a premium, where you pay for the whole node again.
4. DaemonSets you added yourself. Only unmodified system DaemonSets are free. Your logging and security agents are billed on their requests, once for every node they run on.
5. How you will measure it. Enable GKE cost allocation and capture a month of per-team cost on Standard first, so the before and after sit on the same page.

The short version

GKE Autopilot costs $0.0445 per vCPU-hour and $0.0049225 per GiB-hour of pod requests in us-central1, plus $0.10 per cluster-hour, with the first cluster's fee covered by a $74.40 monthly credit. At 500 vCPUs of requests that is about $23,430 a month on demand, $12,890 on a three-year GKE CUD and $7,010 on spot. It beats Standard when your nodes run under about half full, and it punishes oversized requests at a higher rate than Standard does, so rightsizing comes first. Whatever you change, check it on the bill: CostAnalyst connects your Google Cloud billing export read-only and gives each team its Google Cloud cost view, so the Autopilot bill and the savings sit on the page finance reads.