Windows ChatGPT Desktop 26.901.51231 displays raw rich-response markup instead of rendering components

What version of the Codex App are you using (From “About Codex” dialog)?

Desktop 26.901.51231

What subscription do you have?

Plus

What platform is your computer?

Microsoft Windows NT 10.0.19045.0 x64

What issue are you seeing?

The ChatGPT section of the combined ChatGPT/Codex Windows desktop application is displaying raw internal response markup instead of rendering rich content. This began recently, after the application was working correctly a few days earlier.
When I ask a question that produces a response with web citations, images, links, layout elements, or other rich content, the response contains visible component tags and raw reference identifiers. Examples include box/row layout tags, AsyncImageGroup, Cite, and WritingBlock. In one instance, a writing block displayed only its component name and a caret instead of the text I was supposed to copy. The markup makes the response difficult or impossible to read.
The issue occurs in the ChatGPT section, not Codex. Older messages in the same conversation still render correctly. The exact same conversation renders correctly in the ChatGPT mobile app and in a normal desktop browser, including its images and links. A brand-new chat in the desktop app reproduces the problem with the same prompt.
I have completely exited the application, including the system-tray process, and reopened it. I have also checked for updates, and the application reports that it is up to date. Neither action resolved the issue.
The application version is 26.901.51231, released September 5, 2026. The issue may be related to a recent rendering or response-format compatibility regression, but I have not confirmed the cause.
Feedback ID: no-active-thread-01a07b22-4a13-7201-be2e-e3b98d6dcc5a

Example of failed output from app:

The commonly cited comparison is approximately 18 degrees for a 1911 and 22 degrees for a traditional Glock, measured relative to a perpendicular reference to the slide. Those numbers are useful for understanding the difference, but manufacturers don't always measure grip angle the same way. I could not verify a precise published angle for your HK45C or the SAP-C, so I wouldn't claim they are identical. <Cite ref="turn0search12"/>
<svg viewBox="0 0 340 170" width="100%" xmlns="http://www.w3.org/2000/svg">
  <rect width=340 height=170 rx=12 fill="#F4F4F5"/>
  <text x=170 y=20 fontSize=11 fontWeight="600" textAnchor="middle" fill="#52525B">Conceptual grip-angle comparison</text>
  <g transform="translate(24 38)">
    <rect x=0 y=0 width=125 height=23 rx=3 fill="#71717A"/>
    <rect x=12 y=23 width=34 height=76 rx=5 transform="rotate(-18 29 23)" fill="#A1A1AA"/>
    <text x=62 y=119 fontSize=11 textAnchor="middle" fill="#18181B">1911-style</text>
    <text x=62 y=133 fontSize=9 textAnchor="middle" fill="#52525B">More upright</text>
  </g>
  <g transform="translate(191 38)">
    <rect x=0 y=0 width=125 height=23 rx=3 fill="#71717A"/>
    <rect x=12 y=23 width=34 height=76 rx=5 transform="rotate(-22 29 23)" fill="#A1A1AA"/>
    <text x=62 y=119 fontSize=11 textAnchor="middle" fill="#18181B">Traditional Glock</text>
    <text x=62 y=133 fontSize=9 textAnchor="middle" fill="#52525B">More raked</text>
  </g>
</svg>

<caption>The drawing exaggerates the distinction slightly for clarity; it is not a dimensional drawing of either firearm.</caption>

The practical difference is how the sights naturally align when you present the pistol. Someone accustomed to a 1911 or HK45-style grip may find a traditional Glock points a little high until they adjust their wrist. A more upright grip can feel more natural to that person. Neither angle is inherently more accurate; familiarity, hand shape, and grip geometry matter. <Cite ref="turn0search12"/>

What steps can reproduce the bug?

1. Open the combined ChatGPT/Codex Windows desktop application and select the ChatGPT section.
2. Start a new chat or open an existing conversation.
3. Submit a prompt that requests research involving web sources, images, or citations. For example: “Please research the Strike Arms SAP-C frame and summarize its specifications, independent reviews, and any reported issues.”
4. Allow the response to generate. Rich-content components appear as raw markup instead of rendering as images, links, citations, or formatted text.
5. Open the same conversation in a normal desktop browser or the ChatGPT mobile app. The response renders correctly in those clients.
6. Return to the Windows desktop application. The same response continues to display raw markup.
   The issue is reproducible in both new and existing chats. Completely exiting and restarting the application does not resolve it.

What is the expected behavior?

The expected behavior is that the ChatGPT desktop application renders assistant responses as readable, formatted content rather than exposing internal markup or component definitions.
Images should display as images, citations and links should be usable, and layout elements should render normally. Writing blocks should display their full text and allow the user to copy or edit the content as intended.
The same response should render consistently across the Windows desktop application, the normal web browser, and the mobile app. Raw component tags, internal reference identifiers, and formatting syntax should not be visible to the user.

Additional information

The issue was first observed on September 7, 2026, during my first use of the desktop application since the Sep 5th installed release, and after a recent computer restart. The application had previously rendered rich responses correctly. The About dialog identifies the installed release as September 5, 2026, so the timing may be relevant, although I cannot confirm that the update caused the issue.
The computer recently experienced a BIOS configuration problem that required restoring some BIOS settings. Windows is now operating normally, and the same ChatGPT conversation renders correctly in a standard browser on this computer. I am including this detail for completeness, not because I have evidence that the BIOS issue is related.
This affects ordinary ChatGPT conversations, including product research and general questions, while the ChatGPT section is selected. It is not limited to Codex tasks or a particular project. The problem also makes the application feel substantially less usable because large portions of responses are replaced by raw markup.
I have not reset application data or reinstalled the application. I would prefer to preserve existing Codex project settings and local application state while investigating. I can provide additional diagnostic information or test a proposed fix if needed.

Start by reproducing the issue in the Windows desktop application's ChatGPT section with a new chat and a prompt involving web sources, images, or citations. Compare the same conversation in the browser or mobile app, then investigate the desktop rich-response rendering path. Done means citations, images, links, layouts, and writing blocks render as usable content rather than raw markup.