Codex Cloud (2026): How It Works, Limits, Always-On Options

Do not index

Codex cloud is OpenAI's hosted way to run Codex: you hand it a task from the web, the ChatGPT app, your IDE, or the CLI, it works in an isolated workspace on OpenAI's servers, and it comes back with a diff you can turn into a pull request. OpenAI relaunched it at DevDay on September 29, 2026: you now publish a reusable environment once and every task starts from that prepared filesystem in its own isolated workspace. Cloud tasks start at the Plus plan ($20/mo; Free and Go only get GPT-6 Luna in the desktop app), and cloud work shares your plan's usage allowance with local Codex. If you need Codex on a machine that stays on, keeps its sessions warm, and runs on your schedule, that is still not Codex cloud. That is a persistent instance, which is what we sell at Agent37 Cloud from $4.76/mo, with OpenAI still billing the model usage.

Everything below was checked against OpenAI's current docs on September 30, 2026, the day after the DevDay relaunch. Most explainers still ranking for this term describe the old architecture, which OpenAI now labels Codex Cloud (Legacy) and says it plans to deprecate.

What Codex cloud actually is (and the two things it gets confused with)

OpenAI's docs define Codex cloud as an OpenAI-managed execution environment where Codex can work on repository tasks remotely. A cloud task is a Codex conversation that runs there instead of on your laptop. Since the September 29 relaunch, the flow is: you prepare and publish a cloud environment once (Codex inspects your repos, installs dependencies, and records the setup), then start tasks from desktop, web, or mobile, each in its own isolated workspace, running in parallel while your laptop is closed. The pre-relaunch architecture still exists as Codex Cloud (Legacy) and still powers code review and the GitHub, GitLab (beta), and Linear integrations.

Three different things get called "Codex in the cloud," and they behave very differently:

What people mean	Where the work runs	What stays between tasks
Codex cloud (OpenAI's product)	An isolated workspace on OpenAI's servers, cloned from a cloud environment you publish	The environment's filesystem is reusable across tasks; a task's saved VM state is recoverable for up to 7 days
Codex Remote (phone or second device)	Your own connected Mac or PC, which must stay awake, online, and signed in	Everything, because it is your machine
Codex on a persistent instance (a VPS, or a hosted instance like Agent37)	A computer you control that never turns off	The checkout, caches, installed toolchains, login, and sessions

The first row is what this article is about. The third row is where we come in, and we will be clear about which is which. (A fourth thing now exists too: since September 10, OpenAI's Agents API rents out the Codex harness itself on OpenAI-hosted sandboxes for developers building their own agents. Different product, different article.)

How a Codex cloud task runs, step by step

The relaunched flow has two halves: set up an environment once, then run tasks against it. Knowing the steps explains most of the surprises people hit:

You create a cloud environment by talking to Codex: it checks out the GitHub repositories you select, installs dependencies and tools, and tests the workflow. The result is recorded as an install script plus a start skill you can edit.

You publish the environment. Publishing captures the prepared filesystem, so tasks reuse it instead of rebuilding from scratch, and a background repository refresh keeps checkouts current while preserving dependency caches. Changing the setup later means editing and republishing.

Each task runs in its own isolated workspace cloned from the published environment, on a published VM spec: 2 vCPUs, 8 GiB of memory, and 8 GiB of disk on Plus; 4 vCPUs, 16 GiB, and 32 GiB on Pro, Business, Enterprise, and Edu. At launch there is no separate VM charge.

Internet access follows the environment's setting: Package managers, Custom domains only, or All (unrestricted), with all outbound traffic behind OpenAI's proxy. Secrets are now "network secrets": programs see a placeholder and the proxy substitutes the real value only for allowed HTTPS destinations. Environments can even reach private networks over a configured Tailscale VPN.

The task returns an answer plus a diff you can open a pull request from, and its saved VM state stays recoverable for up to seven days after the last turn, so follow-ups resume instead of restarting.

You can start a task from more places than the web UI. Mentioning @codex on a GitHub pull request starts a cloud chat with the PR as context (on the legacy plumbing) and can push a fix back to the branch. The CLI has a codex cloud subcommand (exec, status, list, apply, diff) that submits work to an environment with --env, --branch, and --attempts for best-of-N runs of up to four; OpenAI's source still labels it experimental. Enterprise workspaces can also hand work to Codex by mentioning @ChatGPT in Slack or Microsoft Teams.

Which ChatGPT plans include Codex cloud, and what it costs

As of September 2026, full Codex starts at Plus: the limited-time promotion that bundled it with Free and Go has ended, and those plans now get exactly one Codex surface, GPT-6 Luna in the desktop app. OpenAI's feature matrix lists Codex cloud on Plus, Pro, Business, and Enterprise/Edu only. Cloud tasks are a plan feature, not a metered product, which is why the pricing question is really a limits question.

Plan	Price	Codex cloud
Free	$0	GPT-6 Luna in the desktop app only (subject to rollout); no CLI, web, or cloud
Go	$8/mo	Same as Free: Luna in the desktop app; full Codex starts at Plus
Plus	$20/mo	Included; local messages and cloud tasks share one five-hour window, plus weekly limits
Pro	$100, $200, or $500/mo	Included; no five-hour limit on Pro currently. Pro 200's allowance drops from 20x to 10x Plus on October 30; Pro 500 adds Ultrafast Astra
Business	$20 per user/mo billed annually (2+ users), $25 monthly	Included; Standard Business shares the Plus five-hour window, with larger VMs for cloud tasks
Enterprise / Edu	Contact sales	Included; flexible-pricing workspaces have no fixed rate limits; without it, per-seat limits match Plus for most features
API key (pay as you go)	Token pricing	No cloud features. CLI, SDK, IDE, and desktop-app local chats

That last row trips up a lot of teams. An API-key sign-in gets you the CLI at token rates, now on the GPT-6 family: GPT-6.1 Sol and GPT-6 Sol are $2 per million input tokens and $10 per million output at short context ($4/$15 past the 272K-token split), GPT-6 Astra is $10/$50 ($20/$75 long), and GPT-6 Luna is $0.10/$0.50 ($0.20/$0.75 long). OpenAI's launch announcement called the 50% cut for Sol and Luna permanent; Astra and GPT-6.1 Sol carry no such pledge, though none of the GPT-6 rates are flagged as promotional the way GPT-5.6 Sol's still are. But cloud tasks, GitHub code review, and the Slack integration are ChatGPT-plan features only.

The allowance math is the part OpenAI publishes least precisely. The pricing docs estimate local messages per five-hour window on Plus and Standard Business at roughly 5 to 45 on GPT-6 Astra, 15 to 160 on GPT-6.1 Sol, 15 to 150 on GPT-6 Sol, and 350 to 3,000 on GPT-6 Luna, and warn these are estimates, not fixed limits. Cloud tasks draw from the same allowance and, in OpenAI's words, "may use more of your allowance than local messages." Hit the limit and Plus and Pro users can buy extra credits instead of upgrading; credits are now token-based with a published rate card, and speed modes bill at multipliers (Fast burns included usage at 2.5x). Your live numbers are at chatgpt.com/codex/settings/usage or /status in the CLI.

The standing rule in the docs today: Plus and Standard Business share the five-hour window, and "Pro plans currently have no five-hour limit," a commitment OpenAI repeated for Pro 200 at DevDay. What changed instead is the Pro lineup. Pro 200 reopened to new subscribers with its allowance cut from 20x to 10x Plus effective October 30 (existing subscribers keep 20x until then, plus a one-time credit grant), and a $500 Pro tier launched above it with Ultrafast Astra access. Plan around the five-hour window on Plus, and around weekly limits everywhere.

Which model Codex cloud uses

This changed with the relaunch: you now pick. Cloud tasks expose a model and reasoning control like local Codex, and since September 29 GPT-6.1 Sol is rolling out as the recommended model on paid plans (it became the CLI's bundled default, while Enterprise and Edu keep it off until an admin enables it), with GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna selectable. The GPT-5.6 family (Sol, Terra, Luna) stays selectable during the rollout, but the clock is running on the old generation: GPT-5.3-Codex-Spark retired on September 14, and GPT-5.5 leaves ChatGPT and Codex on October 14, 2026. Any guide telling you Codex cloud is locked to one automatic model is describing the legacy product.

The Codex cloud limits people actually hit

The marketing line is "delegate and stay in flow." The support forums tell you where that stops:

No access to your machine. Cloud tasks use what is in the cloud environment: your local files, running dev servers, and personal local skills are not transferred. The relaunch did open one door: internal APIs and package registries can be reached over a Tailscale VPN configured in the environment settings. Your laptop still is not on the menu.

Task-scoped, not a computer. The relaunch fixed the old rebuild-everything pain: tasks reuse the published filesystem, and a task's saved VM state is recoverable for up to seven days. But a workspace is still not an always-on box. Nothing runs between tasks, changing the setup means republishing the environment, and anything you want to keep has to land in the diff or the environment.

Internet is restricted by default. The presets are now Package managers, Custom domains only, or All (unrestricted), behind OpenAI's proxy; opening it up is still the security decision the docs warn you about.

Fixed compute specs. OpenAI now publishes them: each Plus task gets 2 vCPUs, 8 GiB of memory, and 8 GiB of disk; Pro, Business, Enterprise, and Edu tasks get 4 vCPUs, 16 GiB, and 32 GiB. Larger or custom VMs are an Enterprise conversation. No per-task timeout is published; at the 2025 launch OpenAI said most tasks take one to thirty minutes.

Quota burn. The pattern in the issue tracker has not changed with GPT-6: one open September report shows a single GPT-6 Astra task in the desktop app consuming 89% of a Plus five-hour window (2.4 million input tokens) in one prompt. Cloud tasks draw from the same pool as everything else you do, and OpenAI says they may draw more.

Scheduling is thin. ChatGPT scheduled tasks can run Codex on a timer, and on Gmail, Slack, and GitHub events on Plus, Pro, Business, Enterprise, and Edu. A task cannot combine an event trigger with a time schedule, desktop scheduled tasks need the app open, and the CLI and IDE have no scheduler at all.

None of that makes Codex cloud bad. It is excellent at a bounded change to a repo you keep on GitHub, running while you do something else, delivered as a PR. The friction starts when the job is long, stateful, or recurring.

Codex cloud vs local Codex vs Codex Remote

ㅤ	Codex cloud	Local Codex (CLI, IDE, app)	Codex Remote
Runs on	An isolated VM workspace on OpenAI's servers	Your machine	Your machine, driven from your phone or another device
Sees	Your GitHub repos plus allowed network destinations	Your filesystem, servers, MCP servers	Same as local
State between tasks	Published environment filesystem; task state recoverable 7 days	Persistent; codex resume picks up sessions	Persistent
Model choice	Pickable; GPT-6.1 Sol is the new recommended model	GPT-6 Astra, 6.1 Sol, Sol, Luna (GPT-5.6 family during rollout)	Same as local
Sign-in	ChatGPT plan only	ChatGPT plan or API key	ChatGPT plan
Needs your laptop on	No	Yes	Yes, awake and signed in
Best for	Parallel, bounded, PR-shaped tasks	Interactive work and anything touching local state	Checking on a running job from the couch

The persistence lives in the middle column, and a computer that stays on is how you get it without keeping a laptop open. OpenAI's remote-connections docs are candid about the trade: the host has to stay "awake, online, and signed in," and they recommend a dedicated always-on computer if you want it to be reliable.

When you want Codex on a computer that stays on

The honest test: if the job fits the task shape, lives in a GitHub repo, and your plan includes cloud tasks, use Codex cloud, you are already paying for it. Reach for a persistent machine when one of these is true:

The work spans days: a migration, a full test suite, a refactor that keeps needing the same warm caches and installed toolchains.

You want it unattended on a schedule: nightly dependency bumps, a code-health report before standup, a cron job that reviews yesterday's PRs.

The agent needs a real network position: a database it can reach, a staging server, an internal API, tools that are not on GitHub.

You are shipping Codex to other people, and each team or customer needs its own isolated box.

Option 1: your own VPS

Any Linux box works. Install the CLI (curl -fsSL https://chatgpt.com/codex/install.sh | sh, currently 0.159.3), run codex login --device-auth because there is no browser on a server (device code login must be enabled in your ChatGPT security settings first), and keep the session alive in tmux. DigitalOcean sells this as a Marketplace 1-click (Ubuntu 24.04, Codex CLI preinstalled, 1 GB minimum) plus a Codex plugin that provisions a Droplet from inside a chat. Three caveats: the 1-click image ships Codex CLI 0.154.0, a few releases behind current, with an update script included; it comes pointed at DigitalOcean's own Serverless Inference by default, whose catalog default is GPT-5.5 (kept or changed on DigitalOcean's schedule, not OpenAI's), so run codex login yourself if you want usage billed to your OpenAI account; and Droplets start at $4/mo for a 512 MB box that is below the image's own minimum, so budget a size up. Prefer a container? Our Codex Docker guide has a tested Dockerfile and the in-container sign-in flows.

Option 2: a persistent instance with Codex preinstalled

Agent37 Cloud is an API that provisions always-on, isolated computers from templates, and one of the system templates is agent37-codex. One POST and the instance comes up with OpenAI's Codex CLI installed:

curl -X POST https://api.agent37.com/v1/instances \
  -H "Authorization: Bearer sk_live_..." \
  -H "Content-Type: application/json" \
  -d '{ "template": "agent37-codex" }'

Then connect your OpenAI account one of two ways. Open the instance terminal (every instance has one at its own URL) and run codex login --device-auth; the login is stored on the instance volume, so it survives restarts and updates. Or, if your app provisions instances headlessly, pass OPENAI_API_KEY as instance env on the create call and the image activates it at boot. From there you talk to Codex over HTTP: POST /v1/responses on the instance URL runs a turn, streams it as server-sent events if you ask, and returns a session_id you reuse to continue the thread. GET /v1/models?agent=codex lists Codex's live model catalog, and model and reasoning_effort are set per turn.

What that buys you over Codex cloud:

The disk persists like a VM. Checkouts, caches, installed packages, and your login stay put across stop, start, and resize.

It runs unattended. Give the instance a cron schedule and the run starts warm; it never needs your laptop, and a sleeping instance bills only its disk.

Real isolation per box. Every instance runs under gVisor with a hard disk quota, on the fleet that runs roughly 1,000 live instances today. SSH, scp, and VS Code Remote SSH work on it.

Bring a different model if you want. Set OPENAI_BASE_URL in the instance env and Codex talks to any Responses-compatible endpoint. Codex is tuned for OpenAI models, so expect rough edges with other families.

It costs less than you would guess. Pricing is one metered rate: $0.80 per vCPU, $0.70 per GB of RAM, and $0.09 per GB of disk per month, billed per minute from a prepaid balance. The 2 vCPU / 4 GB floor shape is $4.76/mo always-on, a sleeping instance bills disk alone, and new workspaces get a $5 starter credit once a card is on file (nothing is charged), roughly a month of the cheapest instance. Full numbers on the pricing page.

And the honest downsides, because a hosting company writing about Codex should say them. Codex is OpenAI's product, not ours: you bring your own OpenAI account or API key, OpenAI bills the model usage, and we host the computer. The agent37-codex image is lean, with no browser and no desktop. Signing in with a ChatGPT plan needs a human in the terminal once for the device code (the fully headless path is an API key). And a 24/7 box does not get you out of OpenAI's usage windows: a Plus login on an always-on instance is still a Plus login, so unattended, high-volume work belongs on an API key.

Setting up Codex cloud in five minutes

If the OpenAI feature is what you need, the current quickstart is short:

Go to chatgpt.com/codex and sign in with your ChatGPT account (Plus or above for cloud).

Connect GitHub and choose the repositories Codex may access. The relaunched environments are GitHub-only for now; GitLab and self-hosted GitHub Enterprise are on the roadmap, and GitLab still works on the legacy flow.

Create a cloud environment by describing your setup: Codex checks out the repos, installs dependencies and tools, records the install script and start skill, and you publish the result. Edit and republish when the setup changes.

Describe the result you want and let it run in the background; you can start and continue tasks from desktop, web, or mobile. From a terminal, codex cloud exec --env <name> --attempts 3 does the same with best-of-three.

Review the diff and open the pull request. Put durable instructions in AGENTS.md rather than in chat; repo-stored skills are available to cloud tasks, personal local skills are not.

Weighing Codex against Anthropic's tool as well? Our Codex vs Claude Code comparison covers models, pricing, and where each wins.

Codex cloud FAQ

Is Codex cloud free?

No, and since September even basic Codex is off the free plans: Free and Go get only GPT-6 Luna in the desktop app, with no cloud features. Treat Plus at $20/mo as the floor for cloud tasks; Pro starts at $100/mo, with $200 and $500 tiers above it.

Can I use Codex cloud with an OpenAI API key?

No. API-key sign-in works for the CLI, SDK, and IDE extension and bills at token rates, but OpenAI's pricing docs list "no cloud-based features" for it. Cloud chats need a ChatGPT plan login.

Does Codex cloud keep my environment between tasks?

Since the September 2026 relaunch, mostly yes: tasks reuse your published environment's prepared filesystem, and a task's saved VM state is recoverable for up to seven days after the last turn. It is still not an always-on computer: nothing runs between tasks, and setup changes require republishing the environment.

Can Codex cloud run on a schedule?

Through ChatGPT scheduled tasks on the plans that include Codex cloud, yes: on a timer, or on Gmail, Slack, and GitHub events. Free and Go get scheduled tasks as a general ChatGPT feature (capped at 3 active) but have no Codex cloud to point them at. Scheduled web runs keep no local folder or worktree between runs, and the CLI and IDE have no scheduler. For a warm, stateful scheduled run you want a persistent instance.

What is the cheapest way to run Codex 24/7?

A small VPS with the CLI in tmux, or a hosted instance. On Agent37 Cloud the agent37-codex template runs on metered compute: the 2 vCPU / 4 GB shape is $4.76/mo always-on, with model usage billed by OpenAI to your own account. New workspaces get a $5 starter credit once a card is on file (nothing is charged), which covers roughly a month of the cheapest instance.

Can I use my ChatGPT plan on an Agent37 instance?

Yes, by running codex login --device-auth once in the instance terminal and approving the code with your ChatGPT account; the login persists on the instance volume. Your plan's usage windows still apply. For unattended provisioning, pass an API key instead.

Does Agent37 sell Codex?

No. Codex is OpenAI's coding agent and needs your OpenAI account. Agent37 Cloud is the infrastructure layer: a persistent, isolated, always-on computer with an API for provisioning, chat, files, and schedules. The same API hosts Claude Code, OpenCode, Hermes, OpenClaw, and any public Docker image.

The bottom line on Codex cloud

Codex cloud earned its September 2026 relaunch: reusable environments, published VM specs, seven-day task state, and a model you finally pick make it the right default for parallel, PR-shaped work on GitHub repos, from the Plus plan up. Its ceiling is still structural: task-scoped workspaces, nothing running between tasks, plan-allowance billing, thin scheduling. When the job outgrows that, put Codex on a computer that stays on. Provision one with the agent37-codex template on Agent37 Cloud: always-on from $4.76/mo, one API call to deploy, and the $5 starter credit (card on file, nothing charged) covers roughly your first month.

Codex Cloud Explained (2026): How It Works, Its Limits, and the Always-On Alternative