What OpenCost does very well
OpenCost is open source under the Apache 2 license, free to use, and a CNCF incubating project. Kubecost originally developed it, and the project FAQ calls it the cost allocation engine of Kubecost. That is a solid base: the code is public and a commercial product relies on it.
Beyond AWS, Azure and GCP, its configuration page lists Oracle Cloud, OVHcloud, STACKIT, Open Telekom Cloud, Alibaba and Scaleway, the last two marked as needing documentation. Its public pages list these capabilities:
- Real-time cost allocation by cluster, node, namespace, controller, service or pod.
- Allocation of CPU, GPU, memory and persistent volumes.
- Prices from the billing APIs of AWS, Azure and GCP, and custom CSV prices for on-prem clusters.
- Cost monitoring of cloud services on AWS, Azure and GCP.
- Export to Prometheus, carbon costs, MCP support, and external costs such as Datadog through plugins.
When OpenCost is enough
Stay with OpenCost if these lines describe you: you avoid a subscription and your data stays in your own systems.
- You run one cluster or a few, and one platform team reads the numbers.
- You already operate Prometheus, with the retention you need.
- You want an open source license and no vendor contract.
- List prices are precise enough for your showback.
- Your team is happy to build dashboards and alerts on the exported metrics.
- Your cost data must stay in your own environment.
What you build around it
OpenCost gives you the measure. A FinOps practice needs more. Your team can do each item below, but each one counts in the real cost of a free tool. Name an owner for each one before you commit.
- Storage. OpenCost requires Prometheus to scrape metrics and store data, so your Prometheus setup decides retention.
- Several clusters. The documentation deploys one instance per cluster, with a central Prometheus-compatible query endpoint such as Thanos, Cortex or Mimir. It warns that labels and cluster IDs must match.
- Accurate amounts. Cloud costs arrive with a delay of several hours up to 24 hours. The documentation says on-demand prices are currently not reconciled with billed costs.
- Access. The install guide opens the interface with a port-forward. To share it, you must expose it and decide who sees what.
- Decisions. The FAQ names recommendations, governance, alerting and federated clusters among what Kubecost adds. Plan your own sizing analysis, alerts and owners.
- Support. OpenCost relies on its community, which you can reach on the CNCF Slack.
OpenCost and Onicost, criterion by criterion
Onicost Kubernetes costs are estimates based on public prices, and they exclude network traffic. With Team, the Cloud space shows the gap between the estimated node cost and the billed amount. The savings shown are also estimates, and no saving is guaranteed.
- Displayed price. OpenCost: free, and you cover its hosting and the time to run it. Onicost: Free up to 32 cores, Team from €39 per month before tax, Business from €99 per month before tax.
- Billing basis. OpenCost: none. Onicost: the cores of your nodes, measured every hour, with no percentage of your cloud spend.
- License. OpenCost: Apache 2. Onicost: a commercial service, with a public Helm chart for its agent.
- Data location. OpenCost: your cluster and your Prometheus. Onicost: France, with Scaleway, in the Paris region.
- Several clusters. OpenCost: one instance per cluster and a central backend to run. Onicost: 10 clusters on Team, no limit on Business.
- Taking action. OpenCost: its pages describe measurement and allocation. Onicost: detailed recommendations on requests, nodes, health and waste with Team, and a scheduler that you enable.
Test both on one cluster
To compare, install Onicost on a cluster where OpenCost already runs. Its guide asks for Kubernetes 1.22 or later, Helm 3.14 or later and outbound HTTPS. It does not list Prometheus: the agent reads usage from the kubelet metrics.
Compare both over the same month. Look at the amounts, then at the hours spent on Prometheus, dashboards, access and alerts. Count those hours at your internal rate, then compare them with the Onicost price for your number of cores. If OpenCost answers your questions for little effort, keep it. If several teams need the numbers and a scheduler, test Team for 14 days with no credit card.
Frequently asked questions
Is OpenCost really free?
Yes. The project states that it is free and open source, under the Apache 2 license. You still pay for the infrastructure that runs it and for your team's time.
Does OpenCost need Prometheus?
Yes. Its install guide says it requires Prometheus to scrape metrics and store data.
Does Onicost need Prometheus?
Its install guide does not ask for it. You need Kubernetes 1.22 or later, Helm 3.14 or later and outbound HTTPS to Onicost.