GKE Cost Check & Autopilot Advisor

It depends on how much of your node capacity Pods actually request. Standard bills every node for every hour, whether Pods reserve 90% of it or 20%. Autopilot bills the CPU, memory and ephemeral storage each scheduled Pod requests, at a higher rate per vCPU than a typical VM. So a well-packed cluster is often cheaper on Standard, and a cluster with a lot of unreserved capacity is often cheaper on Autopilot. This check prices both from your own snapshot and says which side wins under its assumptions, or that the comparison is incomplete.

It runs read-only kubectl commands for nodes and Pods (plus, if allowed, ReplicaSets, StatefulSets, DaemonSets, Jobs, PodDisruptionBudgets, compute classes and the metrics API) and optionally one gcloud describe for node-pool settings. It keeps scheduling and pricing fields only, replaces names with aliases, and never reads Secrets, ConfigMaps, environment variables, images or logs. The file it writes is read in your browser; it is not sent to Finitizer and is never put in the page address.

Autopilot sets a default for containers that request nothing (0.5 vCPU and 2 GiB on the general-purpose class), raises Pods below its minimums, and keeps memory between 1 and 6.5 GiB per vCPU by raising the smaller resource. A Pod asking for 2 vCPU and 16 GiB is billed for about 2.46 vCPU. The comparison includes every such adjustment and lists them.

Some workloads cannot be priced at Autopilot Pod rates: GPU, TPU and machine-series workloads bill per node with a premium, a custom compute class may bill either way, and batch Jobs need a runtime. Subtracting a priced part from the whole Standard bill would flatter Autopilot, so the check shows the priced part and says the comparison is incomplete instead.

Not by itself. Unreserved capacity is allocatable CPU or memory that no Pod request has claimed at capture time. Some of it is availability or scaling headroom on purpose, and PodDisruptionBudgets, anti-affinity, local storage, autoscaler minimums and peak load can all stop a node from draining. The check names the blockers it can see and gives a ceiling in dollars, not a savings figure or a node count.