OpenAI Codex gains reusable cloud environments

OpenAI is giving its Codex coding agent reusable cloud environments that run each task in its own virtual machine and stay available across a developer's devices. The company showed the feature on September 29, 2026 at its DevDay conference.

One virtual machine per task

You point Codex at a GitHub repository; it reads the code, works out the dependencies and writes the install scripts itself. Each task then gets a background virtual machine of its own. Start something on a phone, pick it up in a browser, finish it on a laptop — OpenAI says unsaved work survives the move. Dependency caches refresh in the background without discarding where the task stood.

A task can be picked back up for up to seven days after it was last used.

Who gets it, and how much machine

Access covers ChatGPT Plus and Pro subscribers plus Business, Enterprise, Edu and Healthcare workspaces. Plus accounts land on the smallest virtual machines: half the processors and half the memory of the Pro, Business and Enterprise tiers. Teams can share an environment with agreed settings and permissions, and a shared environment does not expose other people's tasks. OpenAI has not published pricing for the feature.

Network secrets and the Security Cloud

Alongside ordinary environment variables, Codex now supports network secrets: a proxy holds the real value back until traffic reaches an approved domain. OpenAI also launched Codex Security Cloud, which scans GitHub repositories on demand, on a schedule, or automatically on every commit. It collapses duplicate findings and stages fixes in the cloud. The scanner draws on OpenAI's Daybreak Blue cybersecurity models, with no separate application required.

CLI, review, and the gaps

The CLI picks up voice control for kicking off and steering tasks, an /agents view for tracking several jobs at once, and smoother prompt editing, session resume and worktrees. A reworked code review lives in the ChatGPT desktop app, summarizing changes and flagging likely problems even while you are away from the machine; it handles pull requests on GitHub and merge requests on GitLab. The cloud environments themselves reach less far: no GitLab, and no self-hosted GitHub Enterprise Server yet. Computer use and browser use also stay off the table inside those environments.

What we could not verify

Both reports we read trace back to the DevDay announcement. We have no independent numbers on task start-up times, on false positives from the security scans, or on the actual specification of the virtual machines. Whether GitLab review matches GitHub review feature for feature is not stated in either source.