Gpt5.6 sol is AGAIN rerouting to mini instant model

ChatGPT Bugs
1

I would like to ask if anyone else has experienced a similar issue recently.

My ChatGPT interface currently shows GPT-5.6, but the actual reasoning performance seems noticeably different compared with previous usage.

For complex tasks that normally require substantial processing time, such as:

  • reviewing scientific reports;

  • analyzing Excel datasets;

  • checking technical calculations;

  • verifying references and logical consistency;

  • performing multi-step scientific reasoning;

the model now often returns an answer almost immediately.

The concern is not simply that the response is fast. The problem is that some tasks appear to be completed faster than would realistically allow for a thorough analysis.

For example, I uploaded a compressed file containing a Word report and Excel data and requested a comprehensive review, including:

  • checking the report content;

  • verifying the data;

  • evaluating the analysis logic;

  • identifying possible errors.

The response arrived within seconds and claimed to have completed the review, but the depth of analysis did not seem consistent with the complexity of the task.

The observed issues include:

  • The model appears to provide a general evaluation instead of actually checking all details.

  • It sometimes states that it has reviewed files or data without demonstrating file-specific evidence.

  • Complex reasoning tasks feel closer to a quick summary rather than a full analysis.

The model name still shows GPT-5.6, so I am wondering:

  1. Are there different inference modes or resource allocations under the same model name?

  2. Can the displayed model remain the same while the actual reasoning budget changes?

  3. Has anyone else noticed a similar difference in reasoning depth recently?

I am not making a conclusion that the model has changed; I am only reporting a difference in observable behavior and would like to know whether others have experienced the same thing.

2

Sounds like a gpt 5.5 mini routing issue

Try asking the model what model is it

There is a more complex way to confirm it but asking should be alright for beginning

It started happening about a week ago or so was promised to be fixed curious to see if you’re affected again

Are you by any chance using a vpn or using your subscription in a different country to which you bought it in? Are you also on free trial by any chance

4

For the past 10 days, we obviously had a bug with the web application where selecting GPT-5.6 Sol would reroute the request to GPT-5.5 Mini, with absolutely no thinking effort at all.

Now it’s happening again. No matter which model we choose, we’re getting ABSOLUTELY NO THINKING EFFORT — instant responses, no agentic thinking, no real analysis, nothing. And honestly, it’s extremely frustrating.

There’s just one major difference from last time: this time, it actually claims that it is GPT-5.6 Sol.

But you can clearly tell from the behavior and the quality of the responses that something is not right.

For example, the Mini model would typically say something like:

“I’m ChatGPT based on [model]”

while Sol would simply identify itself as:

“I’m GPT-5.6 Sol.”

And here’s an even clearer example:

I gave it the exact same file and task a few days ago. It took around 7 minutes and 30 seconds to properly work through it and produced an acceptable result.

Today, the exact same task took 22 seconds — and the result was completely unusable.

22 seconds vs. 7 minutes and 30 seconds.

That is not a minor difference. That is completely different behavior.

Whatever is happening behind the scenes, something is clearly wrong. If the model is supposed to be using its reasoning and agentic capabilities, we should actually be getting those capabilities instead of an instant response that merely claims to be the selected model.

Come on, OpenAI.

At least let us use the model properly for a week. If there are capacity issues, routing issues, or temporary limitations, just be transparent about them. But silently giving users dramatically different behavior while the interface continues to claim they are using the intended model is incredibly frustrating.

5

Yeah, that monkabrain declared that start thinking instead of really think. It cant analize, cant decide,cant do anything but parroting. itis instant 5.5-mini, not even luna

6

This month , i had 5 days stuck with mini model!!

God know how many days open ai will keeping doing it , its straight rubbery

Even using arena.ai for free would have much better results than this ultra light free model openAI is giving us!

7

This started happening to me yesterday. I even wrote a browser extension to inspect the requests. The interesting thing is that the response is ambiguous, and when I ask about it, it falsely says that it is GPT-5.6 Sol. It answers without actually thinking at all.

It seems that instead of fixing the issue, OpenAI engineers are just hiding more information from users. That is not a good approach. Paying half a month’s subscription fee and then having access to something that can sometimes be used better for free is very frustrating.

I think OpenAI should change its name to CloseAI.

image
image381×580 10.9 KB

8

I’m really impressed! I made this post last time because 5.6 became 5.5 mini: Possible GPT-5.6 Sol routing mismatch: selected GPT-5.6 but server metadata shows GPT-5.5-mini. And then I went back to 5.6 and I didn’t care. What’s happening is GPT is telling me it’s 5.6, but reasoning and thinking is only 5.5! Look at this: GPT-5.6 is shown, but reasoning performance seems significantly reduced?

9

I’m a chatgpt pro member and I haven’t been able to use chatgpt for the past 11 days because of all the issues

10

Its been unusable I got a email from support and they seem aware of the issues but will they fix them i know many other users been having the same problems too

11

Same here, since yesterday. Plus account. Before posting I did a clean test this time: fresh conversation, plain text question, no files, no search, no tools — just to rule out the “mini only ran the tool call” theory. Selected gpt-5-6-thinking, extended effort.

The request itself is fine. Every message event in the stream says gpt-5-6-thinking. Then at the very end of the stream, the server_ste_metadata event quietly says:

model_slug: gpt-5-5-mini
requested_model_experience: thinking
plan_type: plus / paid
tool_invoked: false
turn_use_case: "text"
did_auto_switch_to_reasoning: false
is_autoswitcher_enabled: false
server_ttfvt_ms: 1865

1.9 seconds to first visible token. On an extended thinking request. And when you ask it what it is, it says GPT-5.6 Sol, confidently — so that trick is dead this wave too.

I have another capture from the same day where I did use the python analysis tool. Same result — ste says gpt-5-5-mini there as well (that one took 96s, but that’s the tool running, not the model thinking). So it’s not the tool sub-call. It’s the whole turn.

PoW difficulty was normal on both captures (07a120), so the dirty-IP story doesn’t fit either.

Timeline matches what others are saying: several days of it actually behaving like 5.6 after the fix, then yesterday back to instant mode.

One thing that might save people time: resolved_model_slug is gone from the conversation JSON now, so the old F12 guides don’t work anymore. The executed model only shows up in the live POST /backend-api/f/conversation stream, in the server_ste_metadata event at the very end. That’s where to look now.

Timestamps in case anyone at OpenAI wants to look them up: 2026-09-01T04:56:55Z (tool turn), 2026-09-01T05:05:43Z (text turn).

12

Yeah it’s murky like you can no longer “officially” confirm the model but if you give it complex input and it takes like a sec to produce the result you know something is off

Uh my mate right so he said if you have this issue what helped him personally is a combination of a new account and ip but like a proper ip not vpn but like calling your provider asking for a refresh or just putting airplane mode on your phone for some time, he also mentioned he had to use a new software and hardware fingerprint identity. I don’t know much but he said this helped him. He said it could be related to trial abuse or geo pricing or account sharing like getting a soft nerf due to being a bad actor. Once again it’s my mate so yeah

13

Same ip, same acc for like 5 months straight or so, and almost the most expensive (tax include) plus subs overall and still got this monkabrain 5.5 mini silent reroute. Same as mnt problem. I guess, sonny doesnt like me

14

Same issue here from yesterday evening. Suddenly i got this fast answers and the context was an absolute nightmare now i have to wait and redo some task again as i believe the current model is degraded and cant even understand how my projects are built up, like last week.

This is unacceptable OpenAI.

Interestingly you repair your GPT 5.6 model just before the end of the month so people’s pay again for a month then suddenly the services are back to nothing.

We deserve a final fix finally as now nothing is working and many people losing precious time and money because of you!

15

It’s the Plus plan. I’ve been hit hard…

In my setup, problems started on August 19th, and despite OpenAI’s “resolved” announcement, nothing actually got fixed for several days. Around August 27th, 5.6 finally became usable normally again.

And it was working fine until just 12 hours ago. But just now, the nightmare started all over again. It doesn’t search, it doesn’t think, it spits out answers at lightning speed, and every single one of them is wrong — it’s reverted to being like some ancient bot from years ago…

This keeps happening over and over, so why is it being left unaddressed? If it were free, that’d be one thing, but charging people money for this is unacceptable. What will it take for OpenAI to actually do something about it? Are they leaving it alone because not many people are affected?

16

About an hour ago, it seems OpenAI posted something in a related topic along the lines of “we’re downgrading you to 5.5 as an individual penalty, but we won’t tell you what triggered it.” Are you serious…?

But if that were true, the timing of when this happens should be different for each user. Yet a lot of people have started having the same problem again since around yesterday.

17

Where is this was posted please? I tried VPN then it worked then now back to GPT 5.5 mini again.

data: {“type”:“server_ste_metadata”,“metadata”:{“server_ttfvt_ms”:4543.415785767138,“conduit_prewarmed”:false,“plan_type”:“plus”,“plan_type_bucket”:“paid”,“user_agent”:“web_desktop”,“service”:null,“tool_name”:null,“tool_invoked”:false,“fast_convo”:true,“warmup_state”:“rewarm”,“is_first_turn”:false,“cluster_region”:“westus3”,“model_slug”:“gpt-5-5-mini”,“requested_model_experience”:“thinking”,“product_experience”:“chat”,“region”:null,“is_multimodal”:null,“did_auto_switch_to_reasoning”:false,“auto_switcher_race_winner”:null,“is_autoswitcher_enabled”:false,“is_search”:null,“did_prompt_contain_image”:false,“search_tool_call_count”:null,“search_tool_query_types”:null,“message_id”:“0004fa42-a412-4eed-8d1e-47114bf4a362”,“request_id”:“e9584d0f-11c9-4fdd-801d-4bdbcbfd6fac”,“turn_exchange_id”:“27b102aa-7940-485b-ba48-94dc8861c55c”,“turn_trace_id”:“e440e698-f5b5-4a3c-89bb-4e2b633916b8”,“a32e6ebcb”:1,“resume_with_websockets”:true,“low_turn_topic_ttl”:false,“streaming_async_status”:false,“replace_stream_status”:true,“temporal_conversation_turn”:false,“pro_mode_turn_topic_streaming”:true,“turn_use_case”:“text”,“turn_mode”:“projects”},“conversation_id”:“6a94ab54-0aa0-83eb-9e35-bcf384b2bb3a”}

18

Here is the thread → “5.6 pro model has been automatically downgraded and routed to the 5.5 mini model since its release”

After leaving a comment, OpenAI closed the thread as if to say the matter is closed.
Please read the final comment from OpenAI. With this kind of response, it’s a wonder that paying users haven’t revolted yet.
I now completely understand why those who are highly knowledgeable about AI tend to resent OpenAI…
I must use image2 for my work deadline this month, but I am at a loss for what to do now…

19

Thanks everyone for sharing your reports and sticking with the discussion here.

Our team have confirmed that the team has worked on improving this behavior. Access to certain models or features can still be temporarily limited based on account activity, even with an active paid subscription, and additional details about those checks aren’t available.

Access is reassessed automatically and may return to normal once the activity affecting availability stops. Reviewing the Terms of Use and the model/feature access troubleshooting guidance may also help.

Since remediation has been made and this behavior may still occur in some situations, we’ll close this thread for now. If you continue seeing a prolonged access issue, please work with the Support team through an individual support case at support@openai.com.

Main thread here.

Hope this helps clarify what’s happening.

-Mark G.