OpenAI Codex review

Codex is OpenAI's coding agent. It works in the terminal through the open-source Codex CLI, in VS Code-style editors, Xcode and JetBrains IDEs, in the ChatGPT desktop app, and in the cloud through Codex Cloud ^2,3,4,5,21. It inspects code, makes edits, runs commands and hands back diffs or pull requests ^2,4.

Codex has no subscription of its own. It comes with every ChatGPT plan, shares usage limits with ChatGPT Work, and runs as a mode inside the ChatGPT desktop app ^1,5. OpenAI launched it in May 2025 as a cloud agent inside ChatGPT ²⁸ and upgraded it with GPT-5-Codex that September ²⁹. Today it runs the GPT-6 family, and GPT-5.5 leaves on October 14, 2026 ¹⁴.

Locally, Codex works inside a sandbox the operating system enforces: Seatbelt on macOS, bubblewrap on Linux and WSL2, and a native sandbox on Windows ⁷. In the default Auto preset it reads, edits and runs commands in the working directory, and asks before writing elsewhere or using the network ⁶. Auto-review can pass those approval requests to a separate reviewer agent ⁶. You check edits with /diff or ask for a /review of the working tree, and codex exec runs non-interactively in scripts and CI ⁹.

Codex Cloud runs tasks from a published environment that Codex sets up and tests for your repositories; each task gets its own workspace and keeps going while your computer sleeps ⁴. You review changed files and test results, then commit or open a pull request ⁴. Mention @codex review on a GitHub pull request for a code review ¹⁰, assign a Linear issue to Codex ¹², or ask @ChatGPT in Slack to send repository work to a cloud task ¹¹. Admins set requirements through Agent Security and managed configuration that users can't override ^16,17.

Codex comes with ChatGPT plans rather than its own subscription ¹. Free and Go ($8 a month) get GPT-6 Luna in the desktop app, subject to rollout ¹. Plus at $20 a month adds the CLI, IDE, web and cloud features, and OpenAI pitches it at a few focused sessions a week ¹. Pro costs $100, $200 or $500 a month and currently has no five-hour limit ¹. Business costs $20 per user per month billed annually or $25 monthly, with at least two users; Enterprise and Edu are quote-only ¹. OpenAI publishes ranges per five hours: on Plus, 15 to 160 local messages with GPT-6.1 Sol or 5 to 45 with GPT-6 Astra, and weekly limits may apply on top ¹. Past the limit you spend credits, priced per million tokens, such as 50 input and 250 output credits for GPT-6.1 Sol ¹. Fast mode uses included limits 2.5x faster ¹. For all-day use, budget for Pro at $100 a month.

OpenAI's trust portal lists SOC 2 Type 2 for the API and ChatGPT Enterprise, Edu and Team, and ISO 27001, 27017, 27018, 27701 and 42001 for the API and ChatGPT Enterprise and Edu, plus CSA STAR and PCI DSS ²². Business, Enterprise and API data isn't used for training by default ¹. Enterprise and Edu add data retention and residency controls, EKM, SCIM and role-based access control ¹. Zero data retention and non-US data residency are supported, though under them some safety pauses end a task instead of letting you resume, and local computer access with Work Cloud doesn't provide strict ZDR ^6,20. Locally, network access is off by default and writes stay inside the workspace ⁶.

Independent review volume for Codex on its own is thin. It holds 4.8 from 5 G2 reviews ²⁶ and 4.5 from 48 Gartner Peer Insights reviews ²⁷. Gartner reviewers like its agent workflows and easy tool setup but report drift in long sessions ²⁷, and G2 reviewers say limits run out quickly on Plus ²⁶. On the independent Terminal-Bench 4.0 leaderboard, Codex ties for third at 58.18% with GPT-6 Astra and GPT-6.1 Sol ²⁵. OpenAI names Cisco, Rakuten, Duolingo, Vanta and Instacart as users ²³, and Virgin Atlantic, Ramp and Notion in its April 2026 enterprise announcement ²⁴.