Possible ChatGPT behavior regression after August 19 outage (GPT-5.6 Sol / Projects / long-context workflows)

ChatGPT Bugs
,
1

Hi everyone,

I would like to ask whether anyone else has noticed a change in ChatGPT behavior after the outage that happened on August 19 (around 8 PM ET).

I’m using ChatGPT Plus with Projects and long-running project workflows. Before this incident, the same projects, instructions, and interaction style were working reliably.

After the service recovered, ChatGPT became available again, but the behavior changed noticeably.

This does not look like a normal availability issue. ChatGPT still generates responses, and it often appears to understand the context. The problem is that the way it uses the available context seems different.

The main symptoms I’m seeing:

  • Previously established Project instructions are followed less reliably.
  • The model often appears to understand the task, but switches from execution mode into explanation/planning mode.
  • Instead of producing the requested result, artifact, file, patch, or code change, it may describe what it is going to do, explain the approach, and then fail to actually perform the requested action.
  • It may repeatedly acknowledge the task or restate the plan even when the request is already clear.
  • Long project context appears to be available, but its priority feels reduced compared to before.
  • The model sometimes behaves as if it is answering a new prompt rather than continuing an established workflow.

I also notice the same change in creative writing workflows.

The symptoms there are different in form, but similar in nature:

  • less character-specific behavior;
  • more abstract explanation of emotions instead of showing scenes through actions and reactions;
  • more repeated negative constructions (“not this, not that”);
  • more generic and safe interpretations instead of following established style, character logic, and previously defined rules.

The interesting part is that the model often still knows the relevant information. It can explain the project architecture, the rules, the characters, or the intended approach. However, it does not consistently apply that knowledge during generation.

In other words, the issue does not look like a simple loss of information. It looks more like a change in how context and instructions are prioritized during the actual response generation.

The timing is also notable:

  • before the August 19 outage, the same workflows were working normally;
  • the issue appeared shortly after the outage;
  • the system status later returned to normal, but the behavioral change remained.

The current status page shows ChatGPT as operational, which makes sense because this does not appear to be an availability outage. It feels more like a quality/behavior regression affecting certain workflows.

Environment:

  • ChatGPT Plus
  • Web version
  • Projects enabled
  • GPT-5.6 Sol
  • Windows desktop browser

Has anyone else noticed similar behavior after the August 19 outage?

I would especially like to hear from users who rely on:

  • Projects;
  • long-context conversations;
  • Custom Instructions;
  • complex workflows where preserving previous decisions and context is important.

The key symptom I’m trying to identify is:

The model still seems to know the information, but it no longer reliably applies it.

2

Yes, basically the same. I originally thought it was the fault of the VPN, but even after I closed the VPN, it still could not go back to normal mode.
Did you try asking the Chat about its version? My ChatGPT constantly replied that it is “5.5-mini”/

3

Yes, I’m seeing the same kind of behavior.
I also initially suspected a local issue (browser/VPN/session), but the problem remained after checking those things.

I did try asking ChatGPT about its version. Interestingly, it also sometimes identified itself as “5.5-mini”, even though the model selector in the interface showed a different model.

4

Same problem here. I think it might have something to do with openai limiting chat from employing work capabilities, but if that’s the case they have definitely overdone it, the chat is responding too quickly while using 5.6 sol high effort without seemingly even considering my memory and custom instructions. Very frustrating.

5

Yes. And I’m also seeing very fast loss of attached file context. Files that are initially available and correctly used seem to disappear from the active context after only a few turns. After 4–5 messages, the model may behave as if the attached files are no longer available, even though they are still attached to the conversation.

6

Plus user here, chatgpt suddently not able to write down my character like before, it forgot simple instructions etc…very very annoying!

7

image
image791×728 34.5 KB

Pretty consistent stuff. “Concerning” I think is the right word here.

8

They updated and provided new features especially memory local vs global project memory switching. I had to flip that back and forth on a few projects to get memory and file access to work correctly.

You used up usage allowance on the primary models and you are in fall back mode. Anything that is “mini” is fall back model.

Are you on desktop, website or mobile app?

Try closing out of the app and force close it. If website, logout entirely. clear browser cache. close browser.

It can help to disconnect internet for your device then enable again after a minute.

The last one works for wierd app quirks for me.

If it is on openAi side, none of that helps. Check openai status for outages.

9

Ah. It’s locked on 5.5. mini when asked! ^^"

10

Reading through their article that was updated 7 hours ago, hitting the limit might be the most fitting explanation for now.

However I basically perceived an immediate drop on response quality as I started my day’s use, which makes this limit seem unreasonable to say the very least.

11

Now it seemed that my 5.6 Extra-High could be normally functional, but the 5.6 Pro still thought that it is 5.5 mini.

12

That happens to me when context window fills up and details push off the ledge into nothing.

I have it create summaries and bookmark notes, when I remember to tell it. Never remembers to automatically create them unless you add it in each prompt to check context.

Fall back models hit the forget level faster.

You have a limit on thinking for Plus and Business, I usually use instant. But 5.6 thinking is limited use, supposed instant is not but 160 in/out msgs per 3 hours is easy to exceed when writing or any other serious usage.

I don’t notice too much disruption due to my prep and style of interaction. Typically my long form threads stay reasonably consistent. Periodically I have summaries and details created that I export and upload into my file library.

I moved all my current threads out of projects and use global file library. That helped a lot.

If you just now started experiencing these things after months then they are either updating the models, adding changing features or an outage is going on.

13

Also account type can effect things. I noticed some higher plans have issues plus does not at times.

At least for plus, it has been stable for me. Even with outage reports, I don’t seem to notice most of them.

14

That’s what I suspect openai is tuning towards, limiting chat capabilities so all advanced tasks will be costing work usage. A few minor ui changes occured as well it seems.

15

The model behavior I’m seeing is like this:

User:
“Model, do this, this, and this.”

Model:
“Okay, I’ll do it. I will do this, this, and this. In order to do that, I need to…”
(followed by a long list of steps)

“On the output, the result should be…”
(followed by a detailed explanation of what exactly the result should be, why it is needed, and why this particular approach was chosen)

And then the actual result of the announced and thoroughly explained action is completely missing.

In my opinion, this is not normal behavior, even for GPT-5.5-mini.

This is currently how the model behaves for me with any task more complex than simple interactions like “hello model” or “hi user”.

16

Yeah I’m surprised there aren’t more people talking about this, pretty obvious issue.

17

Yup me too. Is this the end of the honeymoon before terminal enshittification?

18

Im having the same issues with 5.6 a project i normally worked on was working fine now thw model is saying that stuff is missing , its truncated my prompts , its replying wirh verification drafts instead of replying wirh what I’m telling it to do it seems like thinking mode , conversations and the servers are experiencing issues this has paused my workflow because I cant trust the model to work correctly I dont see how chatgpt hasn’t fixed this

19

I’m seeing a very similar issue on Windows desktop app. GPT-5.6 Sol at the highest reasoning setting sometimes responds almost instantly, while the tone/style and use of long-term context feel noticeably different from before August 19.

20

Its happening on the website on mobile , file uploads are having issues, thinking mode is having serious regression, and The Context Drop: The server will silently truncate my massive backstory to save data power, then gaslight me by claiming information is “missing.”

  • The “Verification Draft” Cop-Out: Because the backend cannot allocate enough token processing power to write creative prose under high memory strain, it will lazily spit out summaries or outlines instead of the story.