Codex Annotations Fixed in Our 26.908.70816 Retest

Codex annotations are working again in the site maintainer’s macOS retest of desktop version 26.908.70816 (build 9275), confirmed on September 15, 2026. The previously stalled annotation workflow now responds, the cursor stays visible, and comments can be entered and submitted. The affected workflow has fully recovered in this test. OpenAI’s public release notes do not specifically announce a fix for these symptoms.

Codex annotations: what recovered in this retest

The maintainer confirmed all three practical checks below on the installed build. The version and build number were also read from the local application metadata. This gives readers a concrete version to compare with their own installation.

Previously affected behavior	Result confirmed by the maintainer
Annotation mode became stuck or unresponsive	The annotation workflow responds again
The pointer disappeared while annotating	The cursor remains visible
Comment entry and submission were interrupted	Comments can be typed and submitted

September 15 is the date of this recovery confirmation and this report’s publication. It is not an inferred release date for the binary. We have not identified the first build that changed the behavior, and this macOS result does not establish the status of every Windows report.

For readers who stopped using visual comments because they could not reliably select a target or finish a note, the useful next step is to update the desktop app and retry the same page.

What the earlier missing-cursor problem looked like

A closely matching macOS community report, GitHub issue #36222, was filed on July 30 against Codex Desktop 26.721.81911. Its author described opening a local canvas-based web app and enabling Annotation mode. The normal pointer vanished, while the replacement annotation cursor failed to appear. Leaving Annotation mode restored the pointer.

That distinction matters: an invisible pointer can make selecting an element or drawing a region impractical even when the page itself has not stopped running. The report’s suggested rendering explanation remains the author’s hypothesis; our recovery observation does not prove its underlying cause.

There was also a separate Windows annotation focus report, issue #26325. Pressing Space while writing a comment scrolled the underlying page and disrupted the editor’s focus. It was reported in June on package 26.527.7698.0. That describes interrupted text entry, rather than the macOS pointer disappearing.

These original reports document older symptoms. Neither reviewed issue supplies an official closure or a confirmation that its exact reproduction is fixed in build 9275.

What OpenAI documents about annotation mode

OpenAI’s browser documentation describes the intended sequence: enable Annotation mode, choose an element or area, save a comment, then ask the assistant to address the comments. The built-in desktop browser supports reviewing local web apps alongside the conversation.

The same documentation describes an Adjust control for more precise styling feedback, including spacing, text and color. This is useful when a requested change concerns a particular part of a rendered page.

The public changelog lists the September 11 desktop update under version family 26.908. It includes browser-tab stability improvements, but does not explicitly name the annotation freeze or invisible-cursor fix. The recovery result at the top of this article comes from the maintainer’s test.

Check the desktop version and apply updates

Check the version shown by the desktop application before comparing results. Our confirmed installation is 26.908.70816, build 9275. A Codex CLI version number describes a different component and does not identify the desktop browser build.

Use the desktop app’s Check for Updates command and follow its installation or restart instructions. Once the app has reopened, inspect its version again. An update being available or downloaded does not by itself mean the running application has switched to that build.

OpenAI’s managed-update documentation explains that organizations can disable the built-in updater. In that situation, Settings → General can show In-app updates as Managed, even when Check for Updates remains visible. Ask your workspace administrator about the approved desktop package if this applies to your device.

For the separate question of which product and model your subscription can use, see our GPT-6 access and setup guide.

Repeat a small annotation check on your own page

The following is our suggested verification exercise. It expands the three confirmed outcomes into a repeatable check; it is not a claim that the maintainer tested every page type or keyboard combination.

1. Reopen a page where the problem previously appeared. Start with a visible text element or button, then revisit any canvas or chart that was involved.
2. Enable Annotation mode and move across the page. Confirm that you can see where you are pointing before attempting a selection.
3. Select one element and write a short sentence with spaces. Watch whether typing stays in the editor or unexpectedly affects the page behind it.
4. Save the comment and check that it refers to the intended element. Add another note if that is part of your normal workflow.
5. Submit the comments with a clear request. Confirm they reach the conversation, then inspect the resulting change before accepting it.

Keep the page, window size and browser zoom consistent for the first comparison. Change one condition at a time if the result differs. For an example of checking a rendered application against explicit requirements, use our frontend QA walkthrough.

Make the first recovered comment useful

Once annotation entry works, start with one small change that is easy to verify. Our suggested example is: “This card’s title overlaps its price at this window width. Adjust the spacing so both remain readable, and check the same width afterward.”

That comment identifies a visible defect, the affected elements and an observable result. It gives the assistant a more useful task than “fix the layout.” Review the selected target before sending, particularly when several cards have similar text.

You can then repeat the check after the proposed edit. A working annotation interface confirms that feedback can be delivered; inspecting the page confirms whether the requested change actually worked. Keep those two checks separate when deciding whether to resume a larger review session.