How we compare

Every vendor comparison page eventually turns into marketing. This one tries not to — we tell you exactly what each competitor does well and exactly where we think we're different. If we're wrong about anything, open an issue.

The matrix

	KubeHero	Flexera / Cloudability	Kubecost	OpenCost	CAST AI	PerfectScale	Grafana + Prom only
Attribution accuracy	eBPF 1s resolution	Billing records, 24h stale	cadvisor 5-min avg	cadvisor 5-min avg	proprietary	proprietary	DIY
GPU / TPU first-class	✓ DCGM + MIG + TPU API	Partial	Limited	No	No	No	DIY
Retroactive Savings Plan replay	✓	No	No	No	N/A	N/A	No
Multi-cloud normalization	AWS + GCP + Azure, single schema	Strong	Limited	Limited	AWS-bias	Limited	DIY
Policy enforcement CRDs	BudgetPolicy / CeilingPolicy / RightsizingPolicy	Alerts only	No	No	Proprietary	Proprietary	No
Agentic advisory loop	✓ briefings + actions as guarded CRDs, LLM (your key) or rules	No	No	No	Automation, no briefings	Recommendations only	DIY
humanArm default	✓	N/A	N/A	N/A	Varies	Varies	N/A
Reversible within cooldown	✓ 10-min default	N/A	N/A	N/A	Partial	Partial	N/A
Open-source components	Apache agent + CLI + proto + cost-model	Closed	Core open, adv. features paid	Full CNCF	Closed	Closed	Upstream
Self-host + air-gap	✓ Helm + values.airgap.yaml	Limited	✓	✓	No	No	✓
Audit log export	syslog · webhook · S3 · SIEM	✓	Limited	No	Limited	Limited	DIY
Keyboard-first UX	✓ cmd-K, URL state, actions	No	Limited	N/A	Proprietary	Proprietary	Grafana
Licensing	Open source, self-hosted (Apache 2.0 / BSL 1.1)	Enterprise-seat + platform fee	Free core + Enterprise	Free	% of savings	Enterprise	OSS
Runs on Prometheus stack	✓ ServiceMonitor + PrometheusRule + Grafana ConfigMap	Separate stack	Bundled	Separate	Separate	Separate	Native

Head-to-head, honestly

Flexera / Cloudability

What they get right: Multi-cloud ingest is polished — AWS, Azure, GCP, Oracle, IBM. Enterprise-ready commercial surface. Finance-team-friendly reporting and chargeback workflows.

Where we differ:

• Their Kubernetes attribution runs on billing records + tag-based allocation rules. 24–48h stale by design. Ours is 1-second eBPF-accurate.
• Their GPU visibility is vendor-level (EC2 p4d spend) not pod-level.
• Their policy engine is alerts. No enforcement, no humanArm, no reversible actions.
• Their Savings Plan handling applies forward from the commit date; historical numbers don't restate.

Pick them if: you're a Fortune 500 CFO office needing multi-cloud billing rollup for 20+ accounts and your Kubernetes footprint is small. Pick us if: Kubernetes is >40% of your compute and you want sub-minute accuracy.

Kubecost / Stackwatch

What they get right: Kubernetes-focused from day one. Great allocation model for namespace / label / team attribution. Open core with a credible enterprise offering.

Where we differ:

• Same 5-minute cadvisor averaging ceiling as everyone in that product generation.
• GPU support is there but not first-class. No MIG-slice attribution, no tensor-core utilization.
• No policy enforcement CRDs. Their "Actions" feature is client-side recommendations, not reconciler-driven.
• Savings Plan / CUD handling is billing-based, not replay-based.

Pick them if: you need allocation for a single-cloud shop and are happy with 5-min accuracy. Pick us if: GPUs matter, or you want enforcement teeth.

OpenCost (CNCF sandbox)

What they get right: It's the canonical cost allocation model in the CNCF ecosystem. Free, open, actively maintained. Kubecost's open core is built on it.

Where we differ:

• No UI. It's a library + a server exposing /allocation. You build the product.
• Same cadvisor accuracy as Kubecost.
• No enforcement layer.

Pick them if: you have engineering capacity to build on top. Pick us if: you want a product.

We can ingest OpenCost allocation rules via an importer if you've already built on their label schema — continuity matters.

CAST AI

What they get right: Genuine autoscaling + rightsizing product. Strong on savings delivered. Commercial polish.

Where we differ:

• AWS-first, GCP okay, Azure catch-up.
• Black-box attribution. You trust their number; you can't audit the derivation.
• No CRD surface. Their actions don't survive a CAST AI uninstall — they run in their infrastructure.
• Proprietary policy engine, no humanArm-level safety guarantees exposed.

Pick them if: AWS-only, comfortable with SaaS-owned policy. Pick us if: multi-cloud, or you need every action to live as code in your cluster.

PerfectScale / ScaleOps

What they get right: Rightsizing specifically. Good observation windows, statistical confidence scores.

Where we differ:

• Single-capability. They rightsize well; they don't attribute multi-cloud cost or run budget policies.
• No K8s-native CRD surface for their actions.
• No GPU story.

Pick them if: you just want rightsizing and are happy running multiple FinOps tools. Pick us if: you want one tool.

Grafana + Prometheus alone

What they get right: The actual data plane. Everyone else (including us) runs on top.

Where we differ:

• They're a library, not a product. You need to build attribution, recommendations, and actions yourself. That's a two-year engineering project.

Pick them + us: we use both. KubeHero ships a ServiceMonitor, a PrometheusRule, and three Grafana dashboards so your existing stack is our UI.

The eight moats

Pulled out of the matrix — each is a feature Flexera / Kubecost / OpenCost / CAST AI / PerfectScale structurally cannot ship without rewriting their data plane.

1. eBPF 1s resolution — everyone else averages over minutes
2. Retroactive Savings Plan replay — historical cost restates when a SP kicks in
3. humanArm: true default — enforcement with safety rails, not alerts
4. Multi-cloud cost-per-second normalization — one number across AWS / GCP / Azure SKUs
5. Keyboard-first operator UX — cmd-K, URL state, action buttons everywhere
6. Open source, self-hosted — agent + CLI + proto + cost-model under Apache 2.0; full source you can audit and run yourself
7. Posture + cost correlation — CVE findings ranked by workload $/day (see Posture)
8. Agentic advisory loop with hard guardrails — briefings that explain the cluster and draft actions as CRDs; the AI is read-only, the operator enforces (see Agents). Dashboards show you charts — an advisor tells you what changed and hands you the diff. Autoscalers act without explaining; ours explains and waits for you to arm.

"What about X?"

We'll add rows to the matrix as people ask. If we missed your tool or you think we're wrong about a row, open an issue.