GCP Active Assist and the Recommender it powers are Google Cloud's own optimization advice, generated free from your actual usage and billing data. Active Assist is the umbrella; Recommender is the API and surface that delivers individual recommendations across more than a dozen categories, with cost recommendations among the most valuable. The platform already knows which of your VMs are idle, which are oversized, and where a committed use discount would pay off. The work is not finding the savings, it is prioritizing and applying them without breaking anything.
This article is part of our Google Cloud cluster. For where Active Assist sits among the broader levers, start with our complete guide to Google Cloud cost optimization, the pillar this piece links up to.
The cost recommendations that matter most
Active Assist produces many recommendation types. For cost, these are the ones that move the bill.
| Recommendation | What it finds | Typical action |
|---|---|---|
| Idle VM | Compute Engine VMs with near-zero utilization | Stop or delete |
| VM rightsizing | Machine types larger than the workload needs | Resize to a smaller type |
| Idle persistent disk | Disks not attached to any running VM | Snapshot then delete |
| Idle IP address | Reserved static IPs not in use | Release |
| Committed use discount | Steady spend that would benefit from a commitment | Buy a CUD on the baseline |
| Idle Cloud SQL / images | Unused managed databases and stored images | Delete or downsize |
The idle and rightsizing recommendations are pure waste removal with little risk. The VM rightsizing one connects directly to how to rightsize Compute Engine VMs with Recommender, which goes deeper on the resize workflow.
How Recommender generates the advice
Recommender analyzes historical usage, typically over a recent multi-week window, and models what the resource actually needs versus what it is provisioned for. An idle VM recommendation means CPU and network stayed near zero for the observation period. A rightsizing recommendation means the machine consistently used a fraction of its allocated CPU and memory, and Recommender proposes a specific smaller machine type with the estimated monthly saving attached. Because the advice is grounded in your real telemetry, it is far more reliable than a generic rule, but it is only as good as the observation window, so a workload that is busy only at quarter-end may be flagged idle mid-quarter.
How to act on recommendations safely
Treat recommendations as a prioritized backlog, not a one-click apply. Sort by estimated monthly saving so you work the biggest items first. For idle resources, snapshot disks and confirm with the owning team before deleting, because idle does not always mean abandoned. For rightsizing, apply changes in a maintenance window and watch performance after the resize. For commitment recommendations, do not buy until you have rightsized, because committing to an oversized baseline locks in the waste; see committed use discounts explained. Work through the list on a schedule rather than once, because new recommendations appear continuously as usage changes.
Want every recommendation worked, not just listed?
Our Google Cloud cost audit pulls every Active Assist recommendation across the estate, validates each against the owning team, and applies them in the right order so commitments come after rightsizing. On the performance model you pay only from realized savings. No savings, no fee.
Book a GCP cost audit →Reading recommendations at scale
Clicking through the console works for one project but not for an estate of hundreds. The Recommender API and the BigQuery export of recommendations let you pull every recommendation across every project into one place, rank them by saving, and track which have been actioned. This is how you turn a scattered set of console hints into a managed optimization pipeline. Build it on the same billing export you use for allocation; see Cloud Billing reports and BigQuery billing export.
The limits to know
Active Assist sees what its observation window sees, so periodic workloads can be misjudged; always check the usage pattern before acting on an idle flag. It recommends within Google Cloud's own products, so it will suggest a commitment but not tell you to re-architect a chatty multi-region design or move data to a cheaper store. And it surfaces opportunities but does not apply them, so the value is realized only when someone works the list. Use it as the starting inventory of savings, then layer the architectural and governance work it cannot see.
Recommendation categories, the Recommender API and Active Assist behavior above reflect Google Cloud as of May 2026. Verify current behavior in Google Cloud documentation before acting, as the platform changes.
The Google Cloud Cost Optimization Field Guide includes our recommendation-triage workflow and the queries that pull Recommender output across an estate. It is the downloadable companion to this article.
The short version
Active Assist and Recommender hand you Google Cloud's own cost advice for free: idle VMs and disks, oversized machines, unused IPs, and commitment opportunities, each with an estimated saving. Treat the output as a prioritized backlog, validate before deleting, rightsize before committing, and pull recommendations across the estate through the API rather than clicking project by project. The savings are real but only realized when the list is worked. When you want every recommendation triaged and applied safely, that is exactly what our Google Cloud cost optimization service delivers.