Work Cloud currently fails to start a new session in one existing ChatGPT project. The same minimal prompt works without a project and in other existing projects. Work Local works in every project I tested.
I first noticed the problem when the DevDay presentation began and new features were being rolled out. There was wider service instability that day, so I initially assumed it was temporary load. The failure continued after the status page returned to green. The timing is an observation; the cause is still unknown.
Latest isolation tests: October 4–5, 2026 (UTC+4). Project names below are anonymized.
| Context | Result with a new Work Cloud session |
|---|---|
| No project | Starts and responds |
| Other existing projects tested | Starts and responds |
| Project A (affected) | Fails on the first prompt, before any visible assistant response |
Steps to reproduce
- Open the affected existing project.
- Select Work Cloud and start a new conversation.
- Send
work cloud working test.
Expected: the session starts and responds.
Actual: Hmm...something seems to have gone wrong.
The same project-dependent behavior occurs in the web version and the macOS Desktop app. The web failure was reproduced with both GPT-6.1 Sol / Max and GPT-6 Luna / Light.
Sources comparison and Google Drive test
| Project | Total Sources | Uploaded files | Saved responses | Drive sources |
|---|---|---|---|---|
| A (affected) | 27 | 11 | 15 | 1 |
| B (working control) | 15 | 7 | 8 | 0 |
Both projects use Default memory. The total Sources count includes saved responses and is different from the uploaded-file count.
I temporarily removed only the Google Drive source from Project A, reloaded, and verified 26 Sources / 0 Drive sources. A fresh Work Cloud session still failed. I then restored the exact same file/link, completed a fresh sync, and verified 27 Sources / 1 Drive source. A fresh Cloud session still failed. The underlying Drive file was left intact.
Removing the Drive source did not resolve the observed failure. A project-specific initialization, context, or source-processing problem remains a possibility, but no root cause has been established.
Earlier Support request to test a new Cloud Work task.
@OpenAI_Support Could the Work Cloud team investigate why a fresh session fails only in this project and advise which diagnostic identifiers I should provide privately? Has anyone else reproduced this exact pattern?
The screenshot shows the minimal test after removing the Drive source. Private project/source names, links, account details, and chat history are excluded.

