OpenAI Codex
OpenAI Codex, after the merger
Codex is no longer an app you download. Since 9 July 2026 it is a mode inside the ChatGPT desktop app, alongside Chat and Work. Here is what it actually is now, what it costs, and the usage-pool detail that catches people out.
Reviewed 17 August 2026. Prices are US list, ex-tax, and change often — verify on official pages.
What it is
Not one product — six surfaces
Codex is a brand covering a coding agent that appears in several places. Understanding which surface you are being sold is most of the confusion.
| Surface | Status, August 2026 | What it is for |
|---|---|---|
| ChatGPT desktop app | Generally available | The main graphical experience. Codex is one of three modes — Chat, Work, Codex — since the July 2026 merger. Parallel project chats, git review, worktrees, scheduled tasks. |
| CLI | Generally available | A terminal agent, installed via script, npm or a Windows installer. The surface most professional users actually live in. |
| IDE extension | Generally available | VS Code, VS Code Insiders, Cursor and Windsurf. Xcode and JetBrains integrate natively — you pick Codex as the agent. |
| Cloud | Generally available | Delegated background tasks that keep running with your machine off, and that can open pull requests when finished. |
| Mobile | Generally available since May 2026 | Inside the ChatGPT iOS and Android apps. Start tasks, review output, approve commands remotely. |
| Chrome extension | Generally available | Parallel work across browser tabs. |
Codex also integrates with GitHub, Slack, Linear and Amazon Bedrock, and ships an SDK and a GitHub Action.
Capabilities
What it can actually do
Taken from OpenAI's own developer documentation rather than marketing copy.
Read, edit, run, test
It reads your files, makes edits, runs commands and tests, and inspects the integrated terminal. Cloud tasks run in an environment you configure — you pick the repo, the setup steps and the tools it may use. Tasks run in parallel and in the background, and you can apply the resulting diffs locally.
Open and review pull requests
Connect a GitHub account and Codex can create PRs from its work. On an existing PR, @codex review posts a code review focused on serious issues, and @codex security review runs a deeper analysis — still a research preview. You can configure automatic review of every PR, and customise the rules in an AGENTS.md file.
Three sandbox modes
read-only answers questions but needs approval to change anything. workspace-write reads, edits and runs commands inside the workspace, and asks before touching anything outside it. danger-full-access removes the sandbox entirely, and OpenAI explicitly does not recommend it.
Off by default, everywhere
Network access is disabled in all sandbox modes unless you enable it, with optional domain allowlisting through a proxy. This surprises people whose builds need to fetch dependencies — it is a deliberate choice, not a bug.
Four modes
on-request is the default in version-controlled folders. never stops asking. untrusted auto-runs known-safe reads but prompts on anything that changes state. granular sets policy per category. You can approve from your phone.
Computer Use and more
2026 added worktrees, scheduled tasks, skills, an installable plugin system with a marketplace, a built-in browser with DevTools access, Computer Use for driving macOS and Windows applications, Record & Replay to turn a demonstrated workflow into a reusable skill, and importers for Claude Code, Claude Cowork and Cursor setups.
Pricing
Included on every ChatGPT plan — with a catch
US list prices, checked 17 August 2026.
| Plan | Price | What OpenAI says about Codex on it |
|---|---|---|
| Free | $0 | "Explore Codex capabilities on quick coding tasks." |
| Go | $8/mo | "Use Codex for lightweight coding tasks." |
| Plus | $20/mo | The baseline tier that all published limits are expressed against. |
| Pro | from $100/mo | 5× the Codex usage of Plus. The $200/mo tier gives 20×. |
| Business | $20/user/mo annual, $25 monthly | Team features and shared workspace controls. |
| Enterprise / Edu | Contact sales | Credit-based flexible billing; usage scales with credits rather than fixed rate limits. |
Two things worth knowing before you pick a tier
The usage pool is shared
OpenAI's own wording: usage from Codex, ChatGPT Work, ChatGPT for Excel and Workspace Agents "draws from the same agentic usage and credit pool." Your coding capacity is not ring-fenced. A colleague hammering Excel agents on a shared Business plan consumes the same budget you wanted for builds.
Limits are 5-hour windows, and credits expire
Rate limits reset on rolling five-hour windows, with separate counters for cloud chats and code reviews. Beyond your allowance you buy credits, which expire twelve months after purchase and are non-refundable. OpenAI publishes no dollar-per-credit figure we could find, so modelling heavy usage cost in advance is genuinely difficult.
Free and Go limits
OpenAI publishes rate-limit tables for Plus and Pro but not for Free or Go. Separately, the Codex cloud documentation lists Plus, Pro, Business, Edu and Enterprise — Free and Go are absent, which strongly implies cloud tasks need a paid plan, though we could not find that stated outright.
The honest part
Limitations and complaints
Where Codex genuinely frustrates people.
| Issue | Detail |
|---|---|
| Shared usage pool | Covered above, and the most common source of unexpected throttling. It is documented by OpenAI, but not prominently. |
| Network off by default | Correct for security, awkward in practice. Any task needing to install dependencies fails until you configure access. |
| Linux is preview-only | The Linux desktop preview arrived on 11 August 2026 for Ubuntu, Debian and Fedora, as .deb and .rpm. Users report input-method problems and partial Wayland support. The CLI has been solid on Linux far longer. |
| Rate-limit reset banking expires quietly | OpenAI added bankable limit resets in June 2026. Trade press reported in August 2026 that banked resets silently expire after 30 days with no visible countdown in the interface. That is press reporting, not an OpenAI statement. |
| Task latency | Cloud tasks typically run between one and thirty minutes. Fine for delegated work, wrong for anything interactive. |
| Security history | A vulnerability allowing malicious GitHub branch names to inject commands and exfiltrate auth tokens was reported and patched in March 2026. |
FAQ
Codex FAQ
Is there still a Codex desktop app?
Not a standalone one. On 9 July 2026 OpenAI discontinued the separate Codex app and merged it into the main ChatGPT desktop app, where Codex sits alongside Chat and Work modes. Existing users were migrated automatically with projects and settings preserved.
Do I need ChatGPT Plus to use Codex?
No. OpenAI states Codex is included across ChatGPT plans including Free and Go. However, the Codex cloud documentation lists only Plus, Pro, Business, Edu and Enterprise, so background cloud tasks appear to require a paid plan. Free and Go usage limits are not published.
Does Codex run on Linux?
The CLI has for a long time. The desktop app entered preview on 11 August 2026 for Ubuntu 24.04/26.04, Debian 13 and Fedora 43/44, on x64 and ARM64. Expect rough edges — reported issues include input-method bugs and incomplete Wayland support.
Can Codex open pull requests?
Yes. With a connected GitHub account it creates PRs from its work, and responds to @codex review and @codex security review comments on existing PRs. It needs Codex cloud configured for the repository and push or admin permissions.
Is my code safe?
Codex sandboxes by default and disables network access in every mode unless you enable it. The macOS sandbox uses Seatbelt, Linux uses bwrap and seccomp, and Windows has a native implementation. That said, danger-full-access removes all of it, and OpenAI says plainly not to use it.
Why did I hit a limit when I barely used Codex?
Almost certainly the shared pool. Codex, ChatGPT Work, ChatGPT for Excel and Workspace Agents all draw from the same allowance. Heavy use of any of them reduces what is left for the others.
Codex or Cursor?
Cursor is an editor you work inside; Codex is an agent you delegate to. Many developers run both — Codex has an official Cursor IDE extension, and August 2026 added a migration path for Cursor skills. Cursor is $20/month for Pro, the same as ChatGPT Plus.
Sources
Where these facts came from
Every price and capability on this page traces to an official page or named press report. Checked 17 August 2026.