[ChatGPT Web] Deep Research conversation can remain stuck in research mode after research stops

Surface

ChatGPT Web (Chat), Windows, Microsoft Edge, Simplified Chinese UI.

What issue are you seeing?

After running Deep Research in an existing conversation, the conversation remained stuck in the Deep Research composer state even after the research task had stopped.

The composer continued to show the Deep Research placeholder “获取详细报告” (“Get detailed report”). Sending an ordinary follow-up such as “continue to the next step based on the research above” did not return to normal chat behavior; instead, the thread showed “研究已停止” (“Research stopped”) again.

This made it look as though once Deep Research had been used in that thread, subsequent turns could only continue as research turns rather than normal Chat turns.

Steps to reproduce

1. Open an existing ChatGPT conversation on the web.
2. Select Deep Research from the + tools menu.
3. Run a Deep Research task and let it finish or stop it.
4. Send a normal conversational follow-up that is intended to use the completed research context without starting a new Deep Research job.
5. Observe that the composer may still show “获取详细报告” and the next turn can immediately display “研究已停止” instead of behaving like normal Chat.

Expected behavior

After a Deep Research job finishes or is stopped, the user should be able to return to ordinary Chat in the same conversation without losing the research context.

The composer should clearly reflect whether Deep Research is currently enabled, and the state should not remain implicitly stuck after the research task ends.

Workaround / current status

The issue was resolved in this case by manually opening the + tools menu and toggling Deep Research again, which returned the composer to normal Chat behavior.

So this is no longer blocking, but the state transition was confusing and looked like a thread-level Deep Research lock.

Additional information

• Observed on 2026-08-31.
• ChatGPT Plus.
• The issue occurred in a long-running conversation that had already used Deep Research more than once.
• No private conversation content, project names, or account identifiers are included in this report.

A clearer selected/unselected indicator for Deep Research in the tools menu, plus an automatic return to normal Chat after a research task ends, would likely prevent this failure mode and make recovery more obvious.

No repository files, tests, or entry points are identified in the report. Start by reproducing the Deep Research state transition in ChatGPT Web, including a completed or stopped task and a normal follow-up; done means the composer clearly reflects the selected mode and ordinary Chat behavior resumes without toggling the tools menu.