GitHub Copilot vs Claude Code vs Codex: cloud sessions compared

GitHub Copilot, Claude Code and OpenAI Codex can all run coding tasks in a cloud environment instead of on your own machine. They differ most in how they bill, how big a machine you get and which code hosts they support. Every fact below links to the vendor's own docs, changelog or pricing page. Where a vendor hasn't published something, we say "not published".

Last verified: · plans, prices, billing rates and limits re-read on docs.github.com, github.blog, code.claude.com, claude.com and learn.chatgpt.com that day. The vendors can change any of them. GitHub says billing for Copilot sandbox usage begins in November 2026.

Cloud sessions compared

Cloud sessions compared: GitHub Copilot, Claude Code and OpenAI Codex, checked 30 Oct 2026. Each fact links to the vendor page it comes from.
What you want to knowGitHub CopilotCopilot in the cloud: cloud sandboxesClaude CodeClaude Code on the web: cloud sessionsOpenAI CodexCodex Cloud tasks
Which plans include it

Copilot Student, Pro, Pro+, Max, Business, Enterprise and Copilot access are marked Included. Copilot Free is marked Not included, and Free users can keep using Copilot Chat on GitHub.com and GitHub Mobile. [Plans, Copilot on GitHub]

Generally available on paid Copilot plans and rolling out to users from 28 Oct 2026, subject to your organization's “Copilot in the cloud” policy. [Changelog]

Pro, Max and Team, plus Enterprise users with premium seats or Chat + Claude Code seats. [Quickstart, Cloud docs]

Claude Code is marked No on the Free plan. Organizations with Zero Data Retention, the HIPAA configuration or customer-managed encryption keys can't start cloud sessions. [Pricing, Cloud docs]

The pricing page marks Codex cloud as available on ChatGPT Plus, Pro, Business and Enterprise / Education. The API-key column has no cloud features. [Pricing]

Free and Go are not in that cloud row. Their cards mention the desktop app only. [Pricing]

Price and how usage is billed

Plan prices: Pro $10 a month, Pro+ $39, Max $100, Business $19 per granted seat, Enterprise $39 per granted seat. Student is free for verified students. Copilot access is $0 with GitHub Enterprise Cloud or $10 per user a month as Copilot Standalone. [Plans]

Model use is billed in GitHub AI Credits ($0.01 each). The sandbox is billed separately in compute seconds, GiB-seconds of memory and GiB-months of snapshot storage (rates in the table below the comparison). [AI credits, Copilot on GitHub, Sandbox billing]

Sandbox usage first draws from your GitHub plan's included compute (for example $18 a month on GitHub Pro, $300 on Enterprise Cloud). After that it is billed to your payment method, or blocked if you have none. GitHub says billing for sandbox usage begins in November 2026. [Included usage, Sandbox billing, Changelog]

Pro is $17 a month billed annually ($200 up front) or $20 monthly. Max is from $100 a month. Team Standard seat is $20 a month on annual billing or $25 monthly, and Premium seat is $100 annual or $125 monthly. Enterprise is $20 per seat a month billed annually plus usage at API rates, with a 20-seat minimum. [Pricing]

Each session counts toward your plan's usage limits, and “there's no separate charge for the cloud VM.” Running tasks in parallel uses limits proportionately faster. [Quickstart, Cloud docs]

Limits reset on a rolling five-hour window, with weekly limits on paid plans. After a limit, paid plans can turn on usage credits to keep working at standard API rates. [Pricing]

Plus is $20 a month. Pro starts at $100 a month, with $100, $200 and $500 plans. Business is $20 per user a month billed annually (2+ users) or $25 monthly. Enterprise and Edu: contact sales. [Pricing]

Local messages and cloud chats share your plan's usage allowance, and weekly limits may also apply. The page says cloud tasks may use more of your allowance than local messages. [Pricing]

Plus and Pro users who reach the limit can buy credits. Business, Enterprise and Edu plans with flexible pricing can buy workspace credits. No separate VM or compute charge is listed on the pricing page. [Pricing]

Where it runs

By default in a cloud sandbox: “an isolated Linux environment” that GitHub creates for the session and hosts. A snapshot is saved when it shuts down, so you can resume from another surface. [Sandboxes, Changelog]

Other choices: a GitHub Actions runner, or a custom environment you register yourself. You pick between Cloud and Actions when the session starts. [Copilot on GitHub, Sandboxes]

Sandbox size (vCPUs, memory, disk): not published. [Sandboxes]

A fresh Anthropic-managed VM for each session, running Ubuntu 24.04 on x86_64. Organizations can instead route sessions to a self-hosted environment on their own runners. [Environments, Quickstart]

Size ceilings, which Anthropic says are approximate and may change: 4 vCPUs, 16 GB of RAM and 30 GB of disk. [Environments]

After a few minutes idle the VM pauses with its files saved, and a paused VM can later be reclaimed. [Environments]

Each cloud task runs in its own VM with its own workspace. Default size by plan: Plus 2 vCPUs, 8 GiB memory, 8 GiB disk. Pro, Business and Enterprise 4 vCPUs, 16 GiB memory, 32 GiB disk. [Environments]

Saved VM state is recoverable for up to seven days after you last start a turn or resume the task. Larger VMs are available for Enterprise by contacting OpenAI. [Environments]

Operating system and base image on the current Cloud pages: not published. [Environments, Codex Cloud]

Environment and setup

Cloud sandboxes clone your repositories in, and you can personalize them through a dotfiles repository. They don't support custom compute, setup steps written for GitHub Actions, or self-hosted runners. [Sandboxes, Copilot on GitHub]

The copilot-setup-steps.yml file, larger or self-hosted runners and Windows apply to Actions sessions only. GitHub says that if you rely on them you need the Actions environment. [Actions setup, Changelog]

You can register your own computer or VM as an environment with the Copilot CLI's --environment-host option. [Changelog]

List of tools preinstalled in a sandbox: not published. [Sandboxes]

An environment is a saved configuration of network access, environment variables and a setup script. The script runs as root on Ubuntu 24.04. You can type it in the dialog or load a file of up to 64 KB. [Environments]

Anthropic snapshots the filesystem after setup and reuses it. The cache is rebuilt when you change the script or allowed hosts and after roughly seven days, and it isn't cached if setup takes longer than about five minutes. [Environments]

Preinstalled languages include Python, Node.js, Ruby, PHP, Java, Go, Rust and C/C++, plus Docker, PostgreSQL 16 and Redis 7. Replacing the base image entirely isn't supported yet. [Environments]

You make an environment by choosing GitHub repositories. Codex inspects them, installs dependencies, tests the workflow, and you review the setup and select Publish. [Environments, Codex Cloud]

The tested setup is saved as an Install script and a Start skill. A new task starts from the published environment's prepared filesystem, while an existing task keeps its own files. [Environments]

Environments can be private or shared with your Enterprise workspace. Not yet supported in cloud environments: computer and browser use. [Environments]

Parallel tasks

You can run several sessions in parallel in the cloud. Each session works on one branch at a time and opens one pull request per task. [Sandboxes, Copilot on GitHub]

Number of simultaneous sessions: not published. [Sandboxes, Copilot on GitHub]

Each task gets its own session and its own branch, and you don't wait for one to finish before starting another. [Quickstart, Cloud docs]

No per-user cap is stated. A scheduled-task page mentions an organization-level limit on cloud sessions running at once but gives no number. [Scheduled tasks]

Parallel tasks use your plan's limits faster. [Cloud docs]

Tasks start from a published environment and keep their own working state, with a separate workspace for each task. [Codex Cloud, Environments]

Number of simultaneous tasks: not published. [Codex Cloud, Environments]

Pull request and branch workflow

You need write access to a repository, otherwise the session is read-only. [Copilot on GitHub, Changelog]

The cloud agent can push to one branch only: the pull request's branch if you mention it on an existing PR, otherwise a new copilot/ branch. Branch protections and required checks still apply. [Risks]

Its draft pull requests must be reviewed and merged by a human. Workflows don't run until someone with write access clicks Approve and run, unless you change that. Commits are signed. [Risks]

When Claude reaches a stopping point it pushes its branch to GitHub. You review the diff and can create the PR from claude.ai/code, and the session stays open for more edits. [Quickstart, Cloud docs]

Optional Auto-fix watches a pull request and pushes fixes for CI failures and review comments. It needs the Claude GitHub App. [Quickstart, Cloud docs]

The GitHub proxy rejects branch deletions and pushes of anything that isn't a branch, such as a tag. Branch naming for interactive sessions: not published. Scheduled tasks use claude/-prefixed branches. [Environments, Scheduled tasks]

You review the changed files and test results, ask for follow-ups, then commit or open a pull request when ready. [Codex Cloud, Environments]

Branch naming, push restrictions and merge rules for cloud tasks: not published on the current Cloud pages. [Codex Cloud, Environments]

Internet access and secrets

A firewall is on by default with a recommended allowlist. Organization and repository owners can customize or disable it. It doesn't apply to MCP servers or setup-step processes. [Firewall]

Secrets and variables live in a dedicated Agents section at repository or organization level. In a cloud sandbox only secrets are available, not variables, and values are masked in logs. Names starting COPILOT_MCP_ are for MCP servers only. [Secrets]

Each environment has one of four network levels: None, Trusted (the default), Full or Custom. GitHub, enabled MCP connectors and the Anthropic API bypass the allowlist. [Environments, Cloud docs]

Environment variables can be read by anyone who can use the environment. Network secrets stay outside the VM and a proxy adds them to requests for the hosts you list. GitHub credentials also stay outside the VM. [Environments, Cloud docs]

Per environment you can allow all public internet, allow package-manager domains, or add specific domains. A Tailscale VPN gives access to private networks. OIDC lets a task get short-lived credentials for cloud resources. The default for a new environment: not stated on the current page. [Environments]

Environment variables are passed straight to programs. HTTP secrets are different: the task sees a placeholder and the proxy swaps in the real value for the allowed HTTPS domains. [Environments]

Supported repositories

“Copilot only works with repositories hosted on GitHub.” [Copilot on GitHub]

A sandbox session can span repositories owned by one organization. An Actions session is limited to a single repository. [Copilot on GitHub]

Cloning in Anthropic-hosted environments and pull request creation require GitHub. GitHub Enterprise Server is supported on Team and Enterprise plans. [Cloud docs]

GitLab, Bitbucket and other hosts can be sent as a local bundle, but the session can't push results back to that remote. [Cloud docs]

Private repositories need the Claude GitHub App. The web UI can add several repositories to one session, while claude --cloud works with one repository at a time. [Quickstart, Cloud docs]

You choose GitHub repositories when you create an environment. [Environments, Codex Cloud]

The Cloud environments page lists GitLab and self-hosted GitHub Enterprise Server as not currently supported in cloud environments, and on the roadmap. [Environments]

A separate GitLab page describes a beta, available on all ChatGPT plans, that reviews merge requests. With a project environment it can also run GitLab-triggered coding tasks, including pushing updates to a merge request branch. OpenAI's pages don't explain how this fits with the Cloud environments page. [GitLab beta]

Stated limits

Actions sessions: 59 minutes at most. Cloud sandboxes: “no one-hour limit on total running session time.” [Copilot on GitHub, Changelog]

Chat data is kept 28 days for Actions sessions and for the life of the account for sandbox sessions. [Copilot on GitHub]

Parallel-session cap and sandbox size: not published. [Sandboxes, Copilot on GitHub]

No maximum session length is published. A foreground command waits 2 minutes by default (up to 10), then moves to the background for up to 30 more minutes. [Environments]

Cloud sessions share rate limits with all other Claude usage. A bundled local repository must be under 100 MB. If your organization uses IP allowlisting, hosted cloud sessions fail unless Anthropic exempts them. [Cloud docs]

Usage is limited by five-hour and weekly windows. Pro plans currently have no five-hour limit. The plan estimates on the pricing page are for local messages, and cloud tasks may use more. [Pricing]

Saved VM state lasts up to seven days. Maximum task length and number of simultaneous tasks: not published. [Environments]

Scroll the table sideways on a phone. "Not published" means we didn't find it on the vendor pages listed under Sources. Prices are before tax.

The key differences in plain English

What Copilot sandboxes cost

GitHub now publishes its sandbox rates, so you can price a session. Sandbox compute and memory draw from the compute included in your GitHub plan first, then bill to your payment method. In GitHub's worked example, 600 seconds with 4 GiB of memory costs $0.0216. [Docs: Billing for cloud and local sandboxes]

Copilot cloud sandbox meters, in USD, as GitHub publishes them. The per-hour column is our own arithmetic (3,600 seconds times the rate).
MeterUnitPricePer hour (our arithmetic)What it measures
ComputeCompute second$0.000024$0.0864 for each hourCharged while a session is running. Not charged while it is stopped. [Sandbox billing]
MemoryGiB second$0.000003$0.0108 for each GiB for an hourCharged on memory allocated to the session, not memory in use. [Sandbox billing]
StorageGiB month$0.005—Snapshot of a stopped session, from the stop until it is deleted. [Sandbox billing]
Included compute per month, by GitHub plan. Sandbox compute and memory can draw from it. [Included usage]
GitHub planIncluded compute per month
GitHub Free$12
GitHub Pro$18
GitHub Free for organizations$12
GitHub Team$18
GitHub Enterprise Cloud$300

This is separate from your Copilot plan's AI credits, which pay for model use. See our Copilot Fusion guide for plan prices and credit allowances. [Docs: GitHub Copilot on GitHub.com and GitHub Mobile · Docs: Usage-based billing for individuals]

Which one should you pick?

This section only restates the sourced facts above. It isn't a ranking.

What the vendors don't publish

We didn't find these on the official pages listed under Sources, so we haven't guessed them.

Sources

Every fact on this page comes from the official pages below, opened on .

FAQ

Has GitHub started billing for Copilot cloud sandboxes?

GitHub says billing for cloud sandbox usage begins in November 2026, and the exact day isn't published on the pages we checked. The rates are published: $0.000024 per compute second, $0.000003 per GiB-second of memory and $0.005 per GiB-month of snapshot storage. Usage first draws from the compute included in your GitHub plan. [GitHub changelog: New connected Copilot cloud experience (28 Oct 2026) · Docs: Billing for cloud and local sandboxes · Docs: Product usage included with each plan]

Do Claude Code cloud sessions cost extra?

Anthropic says each session counts toward your plan's usage limits and there is no separate charge for the cloud VM. Cloud sessions share rate limits with the rest of your Claude usage, so running tasks in parallel uses them faster. On paid plans you can turn on usage credits to keep working after a limit, billed at standard API rates. [Claude Code docs: Get started with Claude Code in the cloud · Claude Code docs: Use Claude Code in the cloud · Claude: Plans and pricing]

Which plans can use Codex cloud tasks?

OpenAI's pricing page marks Codex cloud as available on ChatGPT Plus, Pro, Business and Enterprise / Education, and not on an API key. Free and Go aren't in that row, and their cards mention the desktop app only. [OpenAI Codex docs: Pricing]

Can any of them work on GitLab or Bitbucket repositories?

Not fully. Copilot only works with repositories hosted on GitHub. Claude Code can start from a local bundle of a GitLab or Bitbucket repository but can't push back to it. Codex has a GitLab beta that can push updates to a merge request branch when you create a project environment, but its Cloud environments page lists GitLab and self-hosted GitHub Enterprise Server as not currently supported. [Docs: GitHub Copilot on GitHub.com and GitHub Mobile · Claude Code docs: Use Claude Code in the cloud · OpenAI Codex docs: Codex in GitLab (beta) · OpenAI Codex docs: Cloud environments]

How many cloud tasks can I run at once?

All three allow several in parallel and none publishes a number for individual users. Claude Code mentions an organization-level limit on simultaneous cloud sessions without a figure. For Claude Code, parallel tasks also use your plan limits faster. [Docs: About cloud sandboxes for GitHub Copilot · Claude Code docs: Get started with Claude Code in the cloud · Claude Code docs: Automate work with scheduled tasks · OpenAI Codex docs: Codex Cloud]

Do cloud sessions have internet access by default?

Copilot runs behind a firewall that is on by default with a recommended allowlist. Claude Code's default network level is Trusted, which allows listed domains only. For Codex, the current environment page lets you allow all public internet, package-manager domains or specific domains, and doesn't state the default for a new environment. [Docs: Customizing or disabling the firewall for GitHub Copilot · Claude Code docs: Configure cloud environments · OpenAI Codex docs: Cloud environments]

Which one publishes the VM size?

Codex publishes default sizes by plan (Plus: 2 vCPUs, 8 GiB memory, 8 GiB disk; Pro, Business and Enterprise: 4 vCPUs, 16 GiB, 32 GiB). Claude Code publishes approximate ceilings of 4 vCPUs, 16 GB of RAM and 30 GB of disk. GitHub doesn't publish a sandbox size. [OpenAI Codex docs: Cloud environments · Claude Code docs: Configure cloud environments · Docs: About cloud sandboxes for GitHub Copilot]