Codex Cloud: setup, pricing, limits and deployment

Originally published on Lizard (lizard.build).

Codex Cloud runs coding tasks on computers managed by OpenAI. You connect your repositories, prepare a reusable environment, and give Codex a task. It can edit files and run tests while your laptop is asleep. You return to review the changes, ask for fixes, and open a pull request.

For Codex Cloud login, sign in to ChatGPT. On the web or desktop app, choose Work in > Cloud. On mobile, open Codex and select an environment you have already published.

Start with a small change and test results you can check before giving Codex a larger job.

What we tested

On October 17, 2026, we ran the setup and /healthz prompts below with Codex CLI 0.159.2 in a Small Firecracker sandbox on Lizard. The fixture used Node.js 22.23.2 and npm 10.9.8. Codex installed from the lockfile, added the endpoint, and passed both tests: the new HTTP 200 response and the existing 404 response. The project's build command checked JavaScript syntax and passed.

We also confirmed that codex cloud list --limit 5 authenticated and returned exit code 0. We checked the submission command's flags with --help. We did not submit an OpenAI Cloud task or create an environment in its web interface; those steps follow OpenAI's documentation.

What does Codex Cloud do?

Codex can inspect a repository, investigate a bug, change code, and run the project's tools. A cloud environment holds the prepared repositories, dependencies, and access settings. Each new task starts in its own workspace.

This suits work you can describe and verify: fix a failing test, add an API endpoint, or update a dependency and check the build. For a change that needs several rounds of review, continue in the same task so Codex keeps its working files.

OpenAI distinguishes this workflow from remote access to an agent running on your own computer. Cloud tasks use OpenAI's computers and can continue while your machine sleeps. See the Codex Cloud overview.

How to set up Codex Cloud

Use the web or desktop app to create an environment. Have your GitHub repository, its build instructions, and any required development credentials ready.

1. Connect your repository

Choose Work in > Cloud > Select environment > Create environment. Connect GitHub if prompted, select the repositories the project needs, and choose Get started.

For the first setup, pick a repository whose tests you already know how to run. That gives you a result to compare with the cloud setup.

2. Prepare and check the tools

Codex inspects the project, installs dependencies and tools, and checks the workflow. Supply missing details when it asks. Specify the runtime version if your project depends on one.

For a Node.js project, you could ask:

Prepare this repository for development. Use the Node.js version and
package manager specified by the project. Install dependencies from the
lockfile, identify the test and build commands, and run both.

Report each command and its result. If a check fails, show the error and
explain what needs to change before I publish this environment.

The setup can record an Install script for dependencies and a Start skill for starting services and checking readiness. Review those commands, especially if the project needs a database or private package registry.

3. Publish the environment

Review the setup report and resolve failed checks. Save the configuration, then select Publish. Wait for Environment published before starting a task.

Saving stores configuration. Publishing captures the prepared filesystem for new tasks. Sharing controls who can use the environment. These are separate actions in the current cloud environment workflow.

4. Give Codex one testable change

Start a new task from the published environment. Describe the expected behavior and how to check it. For example, adapt this prompt to a web API repository:

Add a GET /healthz endpoint using this project's existing routing style.
It should return HTTP 200 and JSON {"status":"ok"}.
The endpoint must not expose configuration or credentials.

Add a test that checks the status code and response body. Run that test
and the existing test suite. Show the commands, results, and changed files.
Do not deploy or change database schemas.

This prompt defines a small task. An HTTP response alone does not establish that the rest of the application works, so keep the existing suite in the acceptance criteria.

5. Review the result

Read the diff and test output before accepting the change. Check that Codex ran the commands it reports and that the tests cover the requested behavior. Look for skipped checks and unrelated file changes.

Ask for corrections in the same task, then commit the work or open a pull request. Run your normal CI checks before merging. A successful task response is one input to that review.

How environments and saved task files work

An environment supplies the starting setup. A task accumulates its own changes after it starts.

OpenAI documents a default recovery window of up to seven days for saved VM state after the last turn starts or you resume the task. This is a VM recovery window, not a promise of continuous execution or a conversation-history retention period. Commit important work to Git. Saved-state details

For example, suppose you install a missing tool during a task. That fixes the task you are working on. To make the tool part of future tasks, edit the environment, verify the installation, and republish it.

Codex Cloud pricing and usage limits

Cloud chats are part of eligible ChatGPT plans. OpenAI's published US prices, checked October 17, 2026, are:

Free and Go have limited desktop Codex access in the pricing guide. Check cloud eligibility before choosing a plan. An OpenAI API key pays for supported API-based Codex workflows separately; API-key access does not include Codex Cloud features. OpenAI pricing and feature availability

There is no reliable conversion from a subscription price to a fixed number of cloud tasks. Model choice, task length, context, and tool use affect consumption. Local messages and cloud chats share the plan allowance, and ChatGPT Work also shares usage with Codex.

OpenAI's local-message estimates are not cloud-task quotas. Check the usage dashboard for your remaining allowance and reset times. Eligible plans can buy additional credits. Usage-limit details

Give tasks access to services and secrets

A private registry needs both network access and a valid credential. Allowing its hostname only solves the network part.

The current environment settings distinguish two ways to supply values:

With a network secret, programs receive a placeholder. The proxy substitutes the credential for matching requests to allowed destinations over HTTPS on port 443. Configure the exact domains and test a request during setup. Network secrets

For a download failure, inspect the failed URL. A package registry and its download host may need separate allowlist entries. For an authentication failure, check the credential, its scope, and the destination. Keep development credentials limited to the services the task needs.

Codex Cloud limits to check before a task

The cloud environment guide lists these default resources:

Enterprise customers can ask about larger VMs. Check disk use before running builds that download large dependencies or produce large artifacts.

The same guide currently lists computer and browser use, GitLab, and self-hosted GitHub Enterprise Server as unsupported in the new cloud environment experience. Repository skills are available to tasks; personal skills from your local computer do not sync automatically. Check the current limitations when planning a workflow.

If your acceptance test needs browser interaction, arrange a separate browser-capable test environment. A passing unit test cannot show whether a button works in the rendered app.

Why do some guides describe different behavior?

OpenAI still documents Codex Cloud (Legacy) for existing review and integration workflows. Older instructions may describe setup-only secrets or container caching. Use the new cloud environment settings for the workflow above, and check which experience an integration uses before applying legacy instructions. Legacy transition

Codex Cloud vs Codex CLI

The CLI is an interface you can run on your own computer or a remote machine. Codex Cloud supplies an OpenAI-managed workspace. The CLI can also submit tasks to Cloud.

To work with cloud tasks from a terminal, the CLI reference documents codex cloud, codex cloud exec, and codex cloud list.

codex cloud
codex cloud exec --env YOUR_ENVIRONMENT_ID \
  "Run the project's tests and report failures. Do not change files."
codex cloud list --limit 5

Replace YOUR_ENVIRONMENT_ID with an environment available to your signed-in account. The installed CLI's --help lists supported flags. The submission command starts cloud work; check the task's results afterward.

For scripts that run Codex in the machine where the CLI is installed, see our codex exec deployment-check example.

Where Lizard fits after a cloud task

Once you accept a change, your application needs a deployment target. You may also need a browser to test the rendered app before releasing it.

A practical sequence is to review the Codex Cloud diff, merge it through your normal checks, then deploy that commit to Lizard with Lizard CLI. Apps and databases run as separate resources and can keep running after a development sandbox stops.

For a custom agent workflow, Lizard Sandboxes provide a shell, files, and HTTPS previews. The desktop template adds a desktop with a browser. Sandboxes use Firecracker microVMs, with a container fallback for older volumes or snapshots and failed Firecracker starts. Check getInfo().runtime when that distinction matters.

Running Codex CLI there is a workflow on infrastructure you control. You manage its credentials, process lifetime, and saved work, and pay for model usage under the relevant OpenAI terms. Lizard does not add capacity to your Codex Cloud allowance.

Lizard Pro costs $19 per month and includes $19 of usage. Small Sandboxes cost $0.0162 per hour for 2 vCPU, 4 GB RAM, and 5 GB local disk, billed per second while running, including idle time. Usage beyond the included amount is pay as you go. Persistent Volumes cost extra. These are infrastructure charges; OpenAI model usage is separate. Lizard pricing

If you are building a product that creates a workspace for each customer, our Lovable clone tutorial shows a tested Lizard SDK flow for code execution, private previews, and saving project files.

Common setup problems

Follow the setup troubleshooting guide for the failing step. After a setup change, verify it in a new task from the republished environment.