I am reporting repeated incomplete responses in ChatGPT: the interface shows reasoning or search progress, then returns to a completed-response state without delivering the requested final answer or showing an actionable error. Sometimes only an introductory paragraph remains.
Environment and scope
- Product: regular ChatGPT conversations, including project chats and new chats outside projects.
- Clients: I have experienced the issue in the macOS app and on the web.
- Reasoning settings: failures have been observed with Extra High and reproduced with High. These are the labels verified in the interface; the underlying model identifiers have not been independently verified.
- First noticed: the afternoon of October 29, 2026. All times below are UTC+8.
Reproductions and controls
1. Physics Optics project, October 30
I asked for an explanation of a chapter on interference. ChatGPT produced an outline, but after I said “OK, start teaching,” it did not deliver the explanation. The interface showed activity about organizing the material and a processing time of 5 minutes 39 seconds, while the composer was available for another message. The individual conversation had only two user turns.
During an earlier occurrence, I waited over three hours while the interface appeared to remain in a thinking/executing state. This is elapsed waiting time, not verified backend processing time.
2. New regular chat outside the project, October 30
I asked for World War II death totals for the main belligerent countries. ChatGPT immediately supplied an introduction distinguishing military and civilian deaths. It then appeared to continue working for approximately two minutes before stopping. The interface showed “Thought for 2m 11s” and normal response controls, but no country-by-country figures or completed comparison. There was no visible error.
A separate new chat containing only “Hello” received a normal, prompt response. Instant also worked in my experience. This does not establish that the network or every other service component was functioning correctly.
3. Different project, Extra High, October 31
At approximately 10:17, I asked a procurement question. English translation: “Please confirm which components I need to purchase; website links would be ideal.”
By 10:26, the interface showed “Searched 52 websites” and progress summaries about the power supply, audio configuration, sensor connections, interface levels, and purchase list. No final answer appeared. The heading still said “Thinking,” but the normal voice control had replaced the Stop control. The missing answer was also visible when the conversation was reopened through a read-only view.
4. Same procurement conversation and prompt, High, October 31
At approximately 10:33:19, I explicitly selected High and resent the same original prompt. I did not add instructions to shorten the answer or disable search.
The Stop button remained visible while progress summaries updated. At approximately 10:39, after about five to six minutes, the Stop button disappeared, the normal voice control returned, and the interface announced “Response completed.” It showed “Searched 45 websites,” but no final answer and no visible error. The reasoning heading still said “Thinking.”
I did not press Stop, regenerate, refresh, or send a follow-up during this test. Repeating a prompt in an existing conversation changes its context, so this is a practical comparison, not a perfectly isolated experiment.
Separate Pro-mode observation
On the morning of October 31, I selected Pro in both the Mac app and the web interface and asked the same question. The Mac response was unusually fast, showed no visible subagent activity, and was incorrect; the web response showed longer reasoning and was correct. I subsequently deleted the Mac conversation, so that pair cannot now be independently reviewed here. This does not prove a model substitution or fallback. Please treat it separately unless internal evidence links the behaviors.
What needs investigation
Please investigate the actual termination reasons, tool/search outcomes, and whether final answers were generated, saved, and delivered. Why can an incomplete turn appear completed without an error or recovery instruction?
During testing, a connected conversation reader reported “idle/completed” even while the live interface still showed Stop and updating progress. Those reader-level labels therefore cannot establish the backend execution state. Similarly, a lingering “Thinking” label does not prove ongoing computation.
I have reported the issue privately to Support. Is this specific behavior acknowledged, and is there a confirmed workaround or fix? This public report contains no screenshots or private conversation identifiers.
Related community report with similar symptoms, without assuming a shared root cause: GPT-6 responses repeatedly interrupted — HAR evidence attached