Overview
OpenAI Codex is an agent product aimed at software engineering. After DevDay on September 29, 2026, the centre of gravity moved from the local CLI to the cloud: connect a GitHub repository and Codex analyzes the codebase to identify runtimes, dependency versions and toolchains, drafts an install script and a startup skill, then publishes the result as a snapshot once you have tested it. Every later task starts from that snapshot with dependencies already in place.
Tasks run in isolated cloud workspaces. Closing the laptop does not stop them, and state is kept for 7 days from the last activity, during which you can pick up from the web, desktop app, phone or terminal with uncommitted changes intact. On October 9 the desktop app added Composer input prediction, which suggests the next prompt from the conversation and project context and accepts it on Tab, without counting against usage; on October 11 it moved from beta to all Pro users. OpenAI confirmed on October 8 that Codex and ChatGPT Work together reach 40 million active users.
The CLI was refreshed alongside, adding voice control and an /agents view for parallel tasks, while the ChatGPT desktop app gained a code review interface. GitHub review is generally available and GitLab merge requests are in preview.
Key Features
- Reusable cloud environments: Connecting a repository detects runtimes, dependency versions and the toolchain, then drafts an install script and a startup skill. Once tested, the setup is published as a snapshot and every new task boots from it, so dependencies are installed once.
- Isolated workspaces with cross-device resume: Each task runs in its own workspace and keeps going with the laptop closed. State lasts 7 days from the last activity, and you can resume from the web, desktop app, phone or terminal with uncommitted changes carried along.
- Composer input prediction: Shipped in the desktop app on October 9, 2026 and opened to all Pro users on October 11. It suggests the next prompt from the whole conversation and project context, accepts on Tab and stays editable, without extra usage, and can be switched off in settings.
- VM sizes by plan: Plus and Edu Plus get 2 vCPU, 8 GiB of memory and 8 GiB of disk; Pro, Business and Enterprise get 4 vCPU, 16 GiB and 32 GiB. Plus users also share a five-hour usage window between local and cloud tasks.
- Security and network controls: Network secrets scope which HTTPS services a task can reach, and the environment itself narrows where Codex can go. Secrets are issued per task, so long-lived credentials stay out of the repository.
- CLI and review interface: The CLI supports voice control and an /agents view that gathers parallel tasks in one place, with prompt editing, session resume and worktrees. The desktop review interface shows the PR description, changed files, comments and checks, and you can ask Codex about a change before posting feedback.
Use Cases
- Connect a repository to reusable environments so every task starts from a snapshot with dependencies already installed.
- Run long refactors or test backfills in a cloud workspace, close the laptop, and check results from the phone or web the next day.
- Walk a pull request through the desktop review interface, asking Codex why each change was made before posting comments to GitHub.
- Turn on automatic review so Codex completes a first pass in the cloud and people only look at what it flags.
- Track several parallel tasks in the /agents view, edit a prompt and let it continue without opening a terminal per task.
- Use Composer prediction in long sessions to carry a line of thought forward without re-describing the context.
Pros
- Reusable environments boot from snapshots with dependencies preinstalled, so new tasks skip the install step.
- Tasks live in isolated cloud workspaces with 7 days of state, and uncommitted changes survive a device switch.
- Composer predicts the next prompt from full context, stays editable after Tab and adds nothing to usage.
- VM sizes are tiered by plan, so Plus through Enterprise each map to a compute level.
- Network secrets narrow what a task can reach, keeping credentials out of the repository.
- The /agents view collects parallel tasks in one place, and the terminal UI was rebuilt for long sessions.
- GitHub review is generally available, and the desktop app surfaces PR descriptions, changes and checks directly.
Pricing
Available plans are Plus, Pro, Business, Healthcare, Education and Enterprise; API key holders alone do not use cloud environments. Cloud VM sizes follow the plan: Plus and Edu Plus get 2 vCPU, 8 GiB of memory and 8 GiB of disk, while Pro, Business and Enterprise get 4 vCPU, 16 GiB and 32 GiB. Plus users also share a five-hour usage window across local and cloud tasks. Composer input prediction is included in Pro at no extra usage. Check the official site for current pricing.
Summary
Codex spent the year moving the engineering environment into the cloud. Snapshots handle dependency installs, tasks keep running with the laptop closed, state lasts 7 days and resumes across devices, and Composer prediction removes the repeated context descriptions that long sessions used to require. It fits refactors, test backfills and first-pass reviews, leaving people to judge results locally. The 40 million active user figure shows how fast it spread; teams adopting it should check two things, whether the cloud VM size covers their builds and how the review flow connects to existing GitHub practice.
Version History
- Multiple local computers can be attached per task (2026-11-09): On November 9, 2026 Codex began letting multiple local machines attach to the same account, with each task assigned to whichever machine should run it. You can build and test an app on a dedicated machine while continuing to write code on a laptop, so long jobs no longer tie up the local computer.
- Codex 0.165.0 adds a built-in Vertex AI provider and GPT-6.1 Terra (2026-10-28): Codex 0.165.0 adds a built-in Google Cloud Vertex AI provider through model_provider = gcp and lists GPT-6.1 Terra in its model catalogue. It enables memory v2 by default for new configurations, supports importing memory from external agents, and allows changing reasoning effort mid-conversation, alongside repository filters in the session picker and directory grouping in the agent command centre.
- Codex's Composer predictions are now available to all Pro users. (2026-10-11): OpenAI has moved Codex's Composer predictions from beta to available for all Pro users. It can suggest the next prompt based on the entire conversation and project context. After accepting with Tab, it can still be edited. It is included in the Pro plan and does not count toward usage, and can be turned off in settings.
- Codex desktop adds Composer input prediction (2026-10-09): Codex rolled out input prediction in beta for Pro users on desktop, suggesting the next message based on the conversation. Suggestions are accepted with Tab, do not consume usage and can be turned off in settings. Version 0.162.1 shipped the same day, fixing a TUI crash on multi-line async questions and a startup failure caused by mismatched background server settings and CLI defaults.
- Codex and ChatGPT Work reach 40 million active users (2026-10-08): OpenAI confirmed that Codex and ChatGPT Work together reached a record 40 million active users, and that the banked reset has landed on all accounts so credited usage resets are fully in place for paid accounts
- At DevDay 2026 OpenAI introduced reusable Codex cloud environments, with tasks running in isolated cloud workspaces and resumable across devices (2026-09-29):
- OpenAI launches the new Codex Cloud, Agents API opens preview and supports computer use. (2026-09-29):