OpenAI Codex Gets Reusable Cloud Environments and Code Review

What OpenAI shipped

At DevDay 2026, OpenAI gave its Codex software engineering agent its largest update so far. The release has four parts: reusable cloud development environments, a rebuilt command-line tool with voice control, a code review view in the ChatGPT desktop app, and a security scanning product called Codex Security Cloud.2223 Taken together, these changes make Codex less of a terminal helper and more of an ongoing agent that can keep working when the developer's laptop is closed. In practice, this puts OpenAI in direct competition with Anthropic, GitHub and Cursor, which have each spent the autumn building the same kind of product.

Cloud environments that persist

The cloud environments matter most. Codex could already run tasks remotely, but it treated each one as a separate sandbox. Now developers can set up an environment once and reuse it for later tasks.22 That environment can hold repositories, dependencies, tools, permissions and other settings, and developers can reach it from desktop, web or mobile.21 One report adds that developers can move Codex work that started locally into the cloud. The task then no longer depends on the local machine, and changes can be synced back later.28

OpenAI frames this as a team feature. The pitch is a shared workspace with settings and permissions approved in advance, plus faster task starts.22 Coverage agrees on that point. It differs on how much weight to give the claims. One outlet points out that faster starts are something OpenAI intends, not something the announcement measured.27 Another warns that shared environments are effectively security decisions about what an agent can reach. It suggests teams treat them as carefully as CI credentials, and notes that at least one well-known developer reported a rough Codex Cloud experience on launch day.23 Those warnings are worth taking seriously. A reusable environment also means one misconfigured permission gets reused every time.

Cloud environments are available on Plus, Pro, Business, Healthcare, Education and Enterprise plans.2123

A CLI built for running several agents

In the refreshed Codex CLI, developers can start and steer tasks by voice. A new /agents view lets them delegate work and track several agents at once.21 The update also improves prompt editing, session resumption and git worktree support, and adds a cleaner terminal interface for long sessions.22 Release notes for CLI 0.159.0 add an opt-in instant_interrupt setting, which lets new input redirect Codex while it is still working. They also list stronger sandbox protection for .aws credentials.29 One outlet reports that OpenAI open-sourced the Codex harness on GitHub during the same event. That claim did not appear in the other coverage reviewed here.23

The CLI kept changing after DevDay. Codex CLI 0.163.0 shipped on October 15 with steering mid-response and parallel read-only tool calls.31

Code review moves into the desktop app

The new Code Review view in ChatGPT desktop lets developers read summaries, walk through diffs and question Codex about possible problems before they respond to GitHub pull requests or GitLab merge requests.21 Automatic reviews let Codex do a first pass in the cloud while the developer is away.22 Several outlets say the CLI and code review features reach more plans than the cloud environments do. One describes both as available on all plans.2123

The design detail that matters is that a human still sends the feedback. Codex prepares the review, but the developer posts it.23 That puts OpenAI's approach somewhere between GitHub's and Anthropic's, as the sections below show.

Security Cloud and the wider bet

Codex Security Cloud can scan entire GitHub repositories on demand, on a schedule or whenever new commits arrive. Codex then investigates findings, removes duplicates and prepares fixes in the cloud.22 It includes access to models from OpenAI's Daybreak Blue cybersecurity program without a separate application.26 It is limited to eligible Pro, Business, Enterprise and Edu users.29 One summary stresses that the fixes are proposed, not applied automatically.27

The keynote also introduced GPT-6.1 Sol for Codex and a faster model tier. Codex will also run tasks for Dots, OpenAI's new always-on agents, and that usage comes out of the same plan allowance.23 This is the strategic point. Each new place Codex can run is another place OpenAI's broader agent products can get work done.

Anthropic's Claude Code: a fast release pace and its own code review

Anthropic has not stood still. Claude Code went from v2.1.227 to v2.1.305 between August 10 and October 20, with 66 versions in the changelog, most of them fixes.34 By October 27 it had reached v2.1.312.38 Its Code Review product, which reviews an organization's pull requests, became generally available on October 15, and v2.1.304 added per-repository review levels.3440 Anthropic also released Claude Mods on October 1, a layer that lets extensions change Claude Code's behavior from the inside.36 One outlet reports cloud sessions that keep running on Anthropic's servers for Pro and Max users.33

Mid-October showed how closely these tools track each other. OpenCode 2.0, Codex CLI 0.163.0 and Claude Code 2.1.302 all shipped between October 12 and 16. Claude Code's release turned forked subagents on by default.31 One analysis described the October updates across vendors as being about operations, not model quality: worktree isolation, choosing which model each subagent uses, and context budgets.35 That fits the Codex release, which is mostly about workflow plumbing.

GitHub Copilot: the incumbent reviewer goes further

GitHub has gone further than OpenAI in handing review authority to AI. Since September 1, Copilot code review can approve pull requests in public preview, if an administrator turns that on. The setting is off by default, and approvals are dismissed when new commits are pushed.14 On October 2, GitHub let teams request Copilot reviews through its REST and GraphQL APIs. It also confirmed that the more thorough "Balanced" setting became the default on September 28.11 Copilot review is also in public preview for Azure Repos, though only as an advisory reviewer there.15

GitHub's push has had problems. Copilot Code Review had an outage on September 28, which a reverted change fixed.16 GitHub also launched ReviewBench, a benchmark built with Microsoft that ranks Copilot first. GitHub's own team ran the competitors' initial tests, and independent Martian leaderboards rank Copilot fourth or fifth.13 GitHub's own documentation says Copilot review should supplement human review, not replace it.14 Most recently, GitHub's Teams app lets users ask Copilot to turn a notification thread into a pull request.1819

Cursor: the same playbook, a few months earlier

Cursor got to several of these ideas first. Its cloud agents began starting up to three times faster in August thanks to prebuilt environments. In September it added self-hosted machines, Projects for long-running work, and Rollouts and Security Review bots. In October it added iPhone remote control of local agents.1 Projects uses a coordinator agent that splits work across subagents running on cloud machines.8 Seen this way, OpenAI's reusable environments are catching up to Cursor, not opening new ground. Pricing for Cursor runs from $20 to $200 a month.8

The takeaway

What stands out in the Codex launch is how ordinary each feature has become. Persistent cloud workspaces, parallel agent views, automatic first-pass reviews and scheduled security scans now exist in some form at all four of the major vendors.1112234 The vendors differ mainly on how much they trust their own reviewers. GitHub lets Copilot approve pull requests if an admin allows it, while OpenAI keeps a human posting the feedback.1423

Our reading is that OpenAI's advantage is distribution, not novelty. Putting cloud Codex on Plus and code review on lower tiers brings agent workflows to developers who would never pay for a separate tool.23 The open question is reliability. Benchmark fights and launch-day outages suggest no vendor has yet shown that AI review is trustworthy enough for teams to rely on it.131623