Oracle Halved Its Free ARM Cloud. Here's the $0-vs-$5 Kubernetes Math Now.

Oracle gave developers the most generous free server on the internet for years — 4 ARM CPUs and 24 GB of RAM, forever, for nothing — and then spent the summer of 2026 taking half of it back. No blog post. No email. Just a documentation edit on June 15 and a wave of terminated instances on August 18.

So here is the bottom line up front: you can still run an always-on Kubernetes cluster for exactly $0 a month, and it still comes with more raw RAM than anything else free on Earth. But the free plan now survives on a single node Oracle is allowed to reclaim, in a region where new instances routinely fail to provision at all — while the paid alternative is one flat $5 box with none of that. This post works both sides of that math, honestly, so you can decide which side your workload belongs on.

The cut, with dates and numbers​

On June 15, 2026, Oracle reduced the Always Free Ampere A1 allowance from 3,000 OCPU-hours and 18,000 GB-hours per month to 1,500 OCPU-hours and 9,000 GB-hours — the equivalent of running 2 OCPUs and 12 GB of RAM continuously, down from 4 and 24. InfoQ's July 3 report confirmed what the community had already figured out from documentation diffs and shut-down instances: Oracle published no announcement, sent no customer notifications, and did not initially distinguish between free-only and Pay-As-You-Go accounts in the docs.

The confusion that followed is worth remembering, because it set the tone for everything after. PAYG users who had upgraded to get easier provisioning feared their existing 4-OCPU shapes would silently start billing for the overage. On June 22, Oracle support agents reportedly confirmed by email that the new limits apply only to free-tier accounts and PAYG accounts could keep 4 OCPUs and 24 GB at no charge — but as one commenter pointed out, the official wording said "all tenancies," and nobody builds on screenshots of support emails. Oracle's docs now say the 2-OCPU/12-GB equivalence is for "Always Free tenancies," which quietly corroborates the two-tier story without ever announcing it.

Then came the second death. Around August 18, 2026, Oracle began auto-stopping and disabling instances that still sat above the new limit — an enforcement wave, not a grace period. And buried in the policy is the trap that matters most going forward: if an existing over-limit resource is ever terminated, it may not be possible to recreate it above the new limit. Grandfathered capacity is one termination — maintenance, outage, your own mistake — away from being gone permanently.

The worked comparison: Plan $0 vs. Plan ~$5​

Enough history. Here are the two honest ways to keep a small Kubernetes cluster always on in September 2026, summed to totals.

Read that table carefully, because the honest headline cuts against the premise of this post: free wins on raw capacity, roughly 3× on usable RAM. Twelve gigabytes for nothing is still an absurd deal, and no juggling of any other free tier reproduces it. If your only metric is gigabytes per dollar, stop reading and take Plan $0.

But gigabytes per dollar is not the metric that pages you at 3 a.m. The rest of this post prices the three ways Plan $0 can ruin your week — and they are the entire reason Plan $5 exists.

What $0 still buys (the rest of the field)​

Before the taxes, a quick inventory of everywhere else, so nobody accuses this comparison of strawmanning the alternatives:

• Oracle remainder. Beyond the A1 slice: two AMD micro instances (1 OCPU/1 GB each — nearly unusable for tenant workloads after system overhead, but they exist), 200 GB of block storage, 10 TB of monthly egress, and the free OKE Basic control plane. All of it locked to your tenancy's home region.
• Google Cloud e2-micro. One shared-vCPU/1 GB VM, 30 GB of disk, and 1 GB of egress per month in three US regions — and, crucially, no idle-reclaim clause. It is the only mainstream always-free VM that cannot be taken away for being bored. It is also 1 GB of RAM, which rules out Kubernetes entirely.
• AWS. New accounts created since July 15, 2025 get $100 in credits plus up to $100 more for completing onboarding tasks, and the account closes after 6 months or when credits run out. That is a trial, not a home.
• Render free. 750 instance-hours a month, 512 MB RAM on 0.1 CPU, services that sleep after 15 minutes idle with 30–60-second cold starts, and a free Postgres that expires after 30 days. Fine for demos; not always-on infrastructure.
• Fly.io, Railway, Koyeb. Fly's free allowances are retired for new accounts, Railway has no true free tier (Hobby is $5/mo), and Koyeb closed its Starter tier to new signups after the Mistral acquisition in early 2026.

So the real contest is Oracle-vs-Hetzner, exactly as the table above frames it. Now the taxes.

The three taxes on free, priced honestly​

Tax 1: the reclaim rule. Oracle may reclaim any Always Free instance whose 95th-percentile CPU and network and memory (A1 shapes) all sit under 20% for 7 consecutive days. A real Kubernetes node with actual pods usually clears the memory bar — 20% of 12 GB is 2.4 GB resident, and kubelet plus a few workloads hold that easily. But a staging cluster, a node you just drained, or a quiet week trips all three arms of the AND, and then your cluster is gone and you are rebuilding from scratch. Price of the mitigation (keeping real load on the box, automating rebuilds so a reclaim is an annoyance instead of a disaster): a weekend of Terraform or Cluster API work you would not need on a box nobody can take.

Tax 2: the capacity lottery. Provisioning A1 shapes in popular home regions routinely fails with Out of host capacity — a failure mode so chronic that Oracle documents it ("try a different availability domain, or wait a while") and an entire genre of capacity-hunter scripts exists to retry launches until a slot frees up. Note the interaction with Tax 1: a reclaimed instance is not merely work to rebuild, it is work to rebuild at an unknown future time when capacity exists. The known workaround is upgrading the account to PAYG — Always Free usage stays $0 within limits, and capacity gates reportedly loosen — but that means a credit card on file and trusting per-limit metering, which is exactly the billing anxiety the June confusion created.

Tax 3: no grandfathering. As established above, anything you lose, you re-create at 2/12 — and if you are one of the lucky PAYG accounts still holding 4/24, one termination locks you to the new limit forever. This ratchet only moves one direction, and every future Oracle maintenance event is a roll of the dice on your capacity. There is no mitigation. That is the price.

Weigh those against Plan $5's ledger: no reclaim policy, no capacity lottery (Hetzner provisions or it tells you it is out of stock before you architect around it), snapshots and a console that does not disable your account for inactivity. The $60 a year buys determinism, not gigabytes.

Verdict: a decision rule​

Stay on Oracle's free tier if your workload tolerates reprovisioning: learning Kubernetes, CI runners, batch jobs with external state, side projects whose worst outage is a funny story. Doubly so if your account is PAYG-upgraded and still holding 4 OCPUs and 24 GB — that is the single best free server left on the internet, and the rational move is to enjoy it while automating the rebuild for the day it ends.

Pay the flat $5 the moment anything you care about depends on the box staying up: something linked from your README, something with users who are not you, anything stateful you would hate to restore. Sixty dollars a year is less than one hour of almost any developer's time, and the free tier will eventually collect more than an hour from you — in a reclaim rebuild, a capacity wait, or a 2 a.m. discovery that your node is gone.

The summer of 2026 did not kill free-tier Kubernetes. It clarified its price: $0 a month, payable in attention, at unpredictable intervals. For a hobby, that is still a wonderful deal. For anything else, buy the $5 box and spend the attention on your app instead.

Running your own platform on machines you own is exactly what Bex.co is for — push a git repo, get a running HTTPS service on infrastructure nobody can reclaim. Star the repo on GitHub or deploy your first app today.