GPT-5.6 Sol High: intermittent reasoning collapse, missing thinking-time UI, and possible Temporary Chat state regression

ChatGPT Bugs
1

I am reporting an intermittent issue with GPT-5.6 Sol High on ChatGPT Plus involving reasoning execution, the visible “Thought for Xs” indicator, long-context continuation turns, and possibly Temporary Chat state migration.

This is not simply a subjective complaint about response quality. I have observed several distinct, repeatable behaviors over time.

Main symptoms

  • Complex text requests sometimes receive an almost immediate response despite High being selected.

  • These rapid responses often show no “Thought for Xs” indicator.

  • In affected cases, response depth and integration of prior context are often noticeably worse.

  • Explicit instructions such as “reason deeply,” “perform repeated verification,” or “conduct a comprehensive analysis” do not reliably restore extended reasoning.

  • The problem appears especially often during pure-text continuation turns that depend heavily on previous conversation context.

Timeline

The abnormalities first became frequent during a recent trip to Japan, where I repeatedly observed:

  • long reasoning processes being interrupted;

  • unexpectedly short reasoning on complex tasks;

  • significant inconsistency between otherwise comparable turns.

After returning to the United States, the same class of problems continued, so a Japan-specific network issue does not appear sufficient to explain the full pattern.

A later event appears to have been an important turning point:

  1. I started a Temporary Chat.

  2. In that Temporary Chat, the visible thinking-duration indicator had already stopped appearing and responses were becoming unusually fast.

  3. I then saved/converted the Temporary Chat into a regular conversation.

  4. After that transition, rapid text responses with no visible thinking duration became significantly more frequent in normal conversations.

I am not claiming that Temporary Chat conversion caused the problem, only that there is a clear temporal correlation and that the conversion may be relevant to conversation/session state.

Important control: 4m16s reasoning still worked

During the same period, GPT-5.6 Sol successfully performed a complex analysis and visibly showed approximately:

“Thought for 4m 16s”

The response quality was also consistent with normal extended reasoning.

However, after the webpage unexpectedly crashed/refreshed, the answer remained visible but the previously displayed 4m16s thinking-duration indicator disappeared.

This suggests two separate observable issues may exist:

  1. Some turns never show a thinking-duration indicator.

  2. A thinking-duration indicator that was definitely present can disappear after reload.

The 4m16s example also shows that long reasoning has not disappeared completely from the account.

Controlled black-box test

I also ran a three-turn pure-text constraint-satisfaction test with:

  • no web search;

  • no files;

  • no external tools;

  • no developer tools;

  • no network inspection.

Expected solution counts were:

  • Turn 1: 1

  • Turn 2: 4

  • Turn 3: 2

All three answers were correct.

However, none of the three turns displayed “Thought for Xs.”

This means that absence of the thinking-time indicator does not by itself prove incorrect reasoning, but it appears associated with a much faster execution path.

Current hypothesis

I cannot determine the internal cause from the user side, but the observed behavior is consistent with one or more of the following:

  • High / Extended reasoning not being applied consistently across continuation turns;

  • complexity estimation underweighting short prompts that require substantial inherited context;

  • conversation/session state issues after Temporary Chat → regular chat conversion;

  • reasoning-related state not being correctly propagated between turns;

  • thinking-duration UI or metadata not persisting correctly after reload;

  • an intermittent routing or reasoning-effort regression.

These are hypotheses only.

I am not claiming that the account is permanently downgraded or that GPT-5.6 Sol is definitely being routed to a smaller model.

Questions for OpenAI / engineering

Could someone from OpenAI confirm whether any of the following are known issues?

  1. GPT-5.6 Sol High / Extended reasoning failing to engage consistently on continuation turns.

  2. Short continuation prompts receiving insufficient reasoning despite depending on large prior context.

  3. Temporary Chat → regular conversation state-migration issues.

  4. “Thought for Xs” disappearing after a page reload or crash.

  5. Reasoning-related state or UI metadata not persisting correctly between turns or reloads.

I have screenshots showing:

  • normal visible long reasoning;

  • the 4m16s example;

  • the same conversation after the thinking-time indicator disappeared;

  • rapid no-indicator responses;

  • the controlled three-turn test.

I can also provide exact timestamps and additional user-visible reproduction examples if useful.

No private telemetry, network inspection, or internal routing data was used for these observations.

2

ChatGPT has become much harder to use since around August 20

I want to share something I have been experiencing as a ChatGPT Go user.

Since around August 20, 2026, working with ChatGPT has become noticeably more difficult and frustrating for me.

I use ChatGPT for long-term projects that involve very long conversations and many documents. Until recently, this worked reasonably well. I could continue working on a project over time and trust ChatGPT to keep track of what we had already worked on and established.

Now I am constantly running into problems.

ChatGPT often tells me that it cannot see a document that I already uploaded or that is already part of the conversation. Sometimes it seems to have access to only part of a very long conversation. I then have to remind it to look for information that is already there or upload documents again.

This creates another very serious problem: document versions. When I have to upload a document again, I have to explain which version is the current one, even though that had already been established, or even when it is the same document ChatGPT says it cannot read. When a project involves many documents and versions, this can seriously affect the consistency and traceability of the work.

I have also noticed many more problems with following instructions. Things that were clearly established earlier in a project are sometimes forgotten, changed, or contradicted. I often have to correct ChatGPT because it makes assumptions instead of stopping and acknowledging that it needs to check the information.

As a result, I now spend much more time correcting ChatGPT and reminding it about information it already had, instead of actually moving forward with my projects.

Around the same time, another feature I used constantly disappeared: the small navigation markers that used to appear on the right side of long conversations. They allowed me to jump directly to an earlier part of a conversation instead of scrolling through hundreds of messages.

Now navigating a very long conversation has become extremely difficult. Sometimes I spend a long time scrolling and the conversation suddenly jumps or changes position, making it even harder to find earlier information.

I know that some changes to long conversations have recently been announced, so I wanted to share my experience as a user.

I honestly don’t understand what is happening. I am simply reporting that something has clearly changed in my experience, and that this change has made ChatGPT less reliable and much harder to work with for the kind of work I do.

Honestly, I feel that ChatGPT has become much more “stupid” over the past two weeks. Not necessarily because every answer is worse, but because I can no longer rely on it in the same way to keep track of, follow, and work with information that is already part of an ongoing project.

I really hope OpenAI looks into what is happening, because long conversations, continuity, and the ability to work reliably with documents are essential for those of us who use ChatGPT for serious, long-term projects.

3

I have been having the aqme issues for the past two weeks since August 19th and August 20th Why is GPT-5.6 failing long-context tasks? - #5 by mcmillanjoseph18 I think all of these issues are because openai is preparing there infrastructure and servers for astra I hope openai fixes all the issues before they release astra https://openai.com/index/path-to-astra/