Codex Code Review Brings Pull Requests Into ChatGPT Desktop App

What OpenAI shipped

OpenAI used DevDay on September 29, 2026 to add a dedicated pull request review view to the ChatGPT desktop app. Inside Codex, it collects a PR's description, changed files, comments and CI checks on one screen.1118 From there a developer can read summaries, step through diffs, and question Codex about possible problems before sending feedback to a GitHub pull request or a GitLab merge request.18 GitHub support is generally available. GitLab merge requests are in preview.1114

The headline feature has a catch, and it is the part teams most need to understand. The review view and automatic reviews are two separate things. Opening a PR in the app and asking Codex about it does not post anything to GitHub. Codex only comments on its own once someone has wired up the cloud integration.1416

The review view arrived on the same day as Codex Cloud's reusable environments, Codex Security Cloud and interactive plugin extensions.1417 Several write-ups note that this is why Code Review behaves like a full panel and not a chat command.14

Two products under one name

The desktop view is meant for a human reviewer. It includes a personal inbox sorted into "Assigned to me," "Assigned to my team" and "Authored by me." You can pin PRs, mark files as viewed for a specific revision, and move between stacked, dependent PRs.11 Clicking "Review with Codex" opens a new chat. Findings there stay private until the reviewer decides to share them. Nothing gets approved, commented on or merged automatically.1114 One review called it "a reviewer's cockpit, not a review bot," and that description holds up.14

Automatic cloud review is the bot side, and it needs setup. The developer connects the repository to Codex, which requires push or admin permission. Then they turn on Automatic review under "Review code" in Codex settings, and also turn it on under personal preferences with a review trigger.11 Other coverage describes the same requirements: the ChatGPT Codex connector, the right repository permissions, and auto-review switched on. Once that is done, Codex reviews a PR in the cloud before anyone opens it, and its comments show up both on GitHub and in the desktop view.1516 No cloud environment has to be set up.1116 On GitHub, Codex limits itself to P0 and P1 issues.1120 Developers can also call it on a single PR by commenting @codex review.11

A third surface sits before both of these. The /review command in the CLI, IDE extension and app reviews local changes against a base branch, uncommitted work or a single commit, and it does not touch the working tree.1114

Where the reporting diverges: who gets what

The coverage disagrees on plan availability, and the disagreement is mostly about which surface each outlet means.

The Decoder and several guides describe the feature as available on all plans.1816 One detailed guide is more careful. It points out that the ChatGPT release note does not list plan tiers and tells readers to check their own workspace.14 The automatic review documentation is more specific. A guide built on official docs says automatic GitHub review covers Plus, Pro, Business and Enterprise/Edu, and does not support API-key sign-in.20 A separate setup guide also says automatic cloud reviews need a ChatGPT plan and cannot use an API key, while local /review works on every plan, API keys included.11 A hands-on walkthrough says Free and Go users get Codex in the desktop app, but cloud features such as code review start at Plus.13

Our reading: "all plans" is accurate for the interactive review view and for local review. It overstates things for automatic reviews, which depend on a paid ChatGPT plan, the connector and repository permissions. Pricing is also unclear. Automatic reviews reportedly count toward Code Review usage limits20, but the release note gives no per-review cost.14

How the competition handles automated review

OpenAI is entering a crowded field. Each major rival has spent 2026 deciding how much review should be automated and how to charge for it.

GitHub Copilot is the default for teams already on GitHub, and it has shipped review updates almost every month. In April, GitHub said that starting June 1 each Copilot review would be billed in AI credits under usage-based pricing and would also use GitHub Actions minutes on private repositories.1 Lite and Balanced effort levels reached general availability in August.9 Balanced became the default effort on September 28.2 Starting October 2, reviews can also be requested through the REST and GraphQL APIs.2 Copilot already has the per-user automation that Codex users still have to configure. A personal settings page available on every Copilot plan turns on automatic reviews when you create a PR, co-author one, or move one out of draft, with optional reviews on new pushes.5 Copilot also reads AGENTS.md, REVIEW.md and CLAUDE.md files, and it supports agent skills and read-only MCP servers.7310 One comparison estimates Copilot reviews at $0.05 to $1 in AI credits on Lite and $0.25 to $5 on Balanced, plus Actions minutes.14

Anthropic's Claude Code takes the most heavyweight approach. Its managed Code Review sends several agents to examine each PR on Anthropic's infrastructure. A verification step filters out false positives, and the remaining findings are posted as inline comments ranked by severity. The tool never approves or blocks a merge.21 An Owner has to turn it on, and each repository is then set to review once, after every push, or only on request.21 Reported costs are $15 to $25 per review, with reviews taking about twenty minutes.2226 Sources disagree on its status. A summary of Anthropic's October 15 post says Code Review is now generally available on Team and Enterprise with Quick, Standard and Deep settings.28 A vendor guide published the same day still calls it a research preview.29 The Anthropic post is the more authoritative source, but teams should check their own admin console. Anthropic's own numbers are notable: the share of internal PRs getting substantive review comments went from 16% to 54%.2528 It also reports that Claude writes about 90% of Claude Code's own code.28

Cursor's Bugbot dropped its $40-per-seat price for usage billing on May 11, and Cursor puts the average run at $1.00 to $1.50.3135 Critics point out that Bugbot reviews every PR update by default and bills per run, so a branch that goes through several rounds of revision costs more each time.32 Cursor said in June that Bugbot had become about 22% cheaper and much faster.3334 One pricing roundup also reports that SpaceX acquired Cursor in August and that Bugbot is still on sale.34

Why the design choice matters

What stands out in OpenAI's release is the order of priorities. Copilot, Claude Code Review and Bugbot are mainly bots that post to the PR, with dials for how often they run and how deep they go. Codex's new desktop experience starts from the other direction. A human opens the PR, investigates privately with the model, and chooses what to post.14 The automated pass is still there, but it comes second, sits behind a settings page, and only flags the two highest severity levels.11

That looks like a deliberate response to the main complaint about AI reviewers, which is noise. One independent guide estimated that 15–25% of Copilot review comments were wrong, irrelevant or too vague to be useful.8 Keeping findings private until a person checks them, and limiting automatic comments to P0/P1, is a way to protect the signal in PR threads. The tradeoff is that the reviewer has to do the work. The tool helps people review faster. It does not review in their place.

Cost may decide much of this. Anthropic's managed reviews cost tens of dollars each.22 Bugbot runs about a dollar or more each time it triggers.31 Copilot now charges Actions minutes on top of AI credits.1 OpenAI has not published a per-review price for Codex.14 For teams already paying for ChatGPT, that could make Codex the cheapest option to try, until usage limits apply.

The bottom line

The desktop review view is a real upgrade for anyone who reviews code and wants an AI assistant without a bot posting on their behalf. If you want Codex to post comments automatically, you still need a paid plan, a connected repository, the right permissions and two toggles turned on. A sensible workflow, which one guide also suggests, is to run local /review before pushing, let automatic review take a first pass when the PR opens, and then do your own review in the app.11 Teams comparing tools should look past the launch messaging and check the cost per review and how much noise each one adds to PR threads, because the four vendors differ most on those two points.