Codex Cloud Pricing: $20 Plus, 2 vCPUs and Three Gaps

Codex Cloud runs coding tasks on OpenAI's computers while your laptop sleeps, and the question buyers keep asking is what the subscription actually buys. A setup guide published this week, which tested the workflow with Codex CLI 0.159.2 on October 17, 2026, lays out the prices, the VM sizes and the gaps.

What it costs

Cloud chats are bundled into eligible ChatGPT plans rather than sold separately. The guide lists OpenAI's published US prices, checked October 17, 2026: Plus at $20 per month, Pro at $100, $200 or $500 per month, and Business at $20 per user per month billed annually for two or more users, or $25 billed monthly. Enterprise and Edu are contact-sales. Free and Go are described as having limited desktop Codex access.

The important caveat is that a price does not convert into a task count. As the guide puts it, there is no reliable conversion from a subscription price to a fixed number of cloud tasks — model choice, task length, context and tool use all move consumption, and local messages share the same plan allowance as cloud chats. An OpenAI API key does not unlock Codex Cloud; it pays for API-based workflows separately.

The machine you get

Default resources split along plan lines. Plus and Edu Plus get 2 vCPUs, 8 GiB memory and 8 GiB disk. Pro, Business, Enterprise, Edu and Edu Pro get 4 vCPUs, 16 GiB memory and 32 GiB disk, with larger VMs available to Enterprise on request. On a Plus plan, 8 GiB of disk is the number to check first if your build downloads heavy dependencies or emits large artifacts.

Saved VM state has a documented default recovery window of up to seven days after the last turn starts or you resume a task — a recovery window, the guide stresses, not a promise of continuous execution or history retention. Commit anything that matters to Git.

What it will not do yet

The cloud environment guide currently lists computer and browser use, GitLab, and self-hosted GitHub Enterprise Server as unsupported in the new experience. Repository skills are available to tasks, but personal skills from your local machine do not sync automatically. If your acceptance test needs a browser, you need a second, browser-capable environment — a passing unit test cannot show whether a button works in the rendered app.

Credentials have their own split. Environment variables hand a program the real value; network secrets give it a placeholder and let the proxy substitute the credential for matching HTTPS requests to allowed destinations on port 443. A private registry needs both the hostname allowed and a valid credential, and a package registry may need its download host allowlisted separately.

Worth it, for which job

The honest fit is bounded, verifiable work: fix a failing test, add an endpoint, bump a dependency and check the build — jobs you can describe and then check. The guide's own test was exactly that shape, with Codex installing from a lockfile and passing both a new HTTP 200 test and an existing 404 test on a Node.js 22.23.2 fixture. For terminal users, codex cloud, codex cloud exec and codex cloud list submit and inspect tasks without leaving the shell.

Two caveats before you upgrade a plan for this. The guide is published by a sandbox vendor whose own product appears in the recommended workflow, so read the comparisons with that in mind. And OpenAI still documents a legacy Codex Cloud for older review and integration workflows — check which experience an integration targets before following any older instructions.