Solid evidence: GPT-5.5 caught with "diminished intelligence", acknowledged in official OpenAI documentation.

[Introduction] GPT-5.5 was exposed for "fake thinking". It was secretly replaced with the mini version in just two hours. For a monthly fee of $200, you get a "Schrödinger's brain". The Trace command confirms it, and the official documentation acknowledges it. Later, there were numerous complaints: OpenAI, who are you kidding?

ChatGPT has been reported to have "dumbed down" again!

Just in the past few days, there was an uproar on X first.

Netizen Lisan al Gaib found that after using GPT-5.5 for an hour or two, it suddenly became stupid. Each request was answered instantly, and the quality dropped precipitously.

However, what was still shown on the interface was "GPT-5.5 Extended Thinking".

That is to say, the label of thinking was still there, but the thinking itself had disappeared.

$200 per month, bought a "Schrödinger's model"

On the OpenAI developers' forum, a complaint post also went viral.

Agentify.sh said that GPT-5.5 would suddenly lose the ability to follow instructions while being used.

It excitedly announced that it had "fixed" the problem, but the code quality was so poor that it caused a large - scale rollback.

Previously, UI tasks that could be easily handled by 5.5 - med can't even handle the simplest changes now.

Upgrading to 5.5 - high didn't work. Upgrading to xhigh still didn't work.

Moreover, the time that xhigh could run before has now been significantly shortened.

As soon as the post came out, the reply section exploded.

Some people directly reverted to 5.4.

Some people were using the xhigh top - grade version, but "compared with last week, it's obviously worse. Long tasks frequently make mistakes and it completely doesn't follow the workflow".

Some people reported even more outrageous situations. "Even a simple query takes a long time to process. If you interrupt it to correct the direction, it directly ignores you and continues with the previous wrong plan."

Yes, everyone is describing the same phenomenon - GPT's "brain" has been secretly replaced without knowing when.

To be honest, GPT - 5.5's current performance is similar to that of 5.3. It was amazing in the first few days, but now there's no sign of the original model at all.

It's not an illusion. OpenAI has it in black and white

To verify, Lisan al Gaib specifically conducted a comparative test.

With the same account, the results from the Extended Thinking mode on the ChatGPT side were all garbage. But when switching to the xhigh mode on the Codex side, it immediately returned to normal.

In his own words, Codex is "literally 4 billion times smarter than this thing".

Developer Andrew Curran came up with a clever idea - directly ask the model "What is the cut - off date of your training data?"

The model replied, August 2025.

The problem is that the cut - off date for GPT - 5.5 Thinking is December. August is the cut - off date for the Instant version!

That is to say, he selected Thinking, but the system actually ran the Instant version for him.

The model label on the interface didn't change a single word, but the model behind it had been secretly replaced...

Interestingly, this time OpenAI actually confirmed it in its own help document.

According to the official description on the OpenAI Help Center, Plus users can send a maximum of 160 GPT - 5.5 messages every 3 hours.

After using up the quota, the system will silently switch to the mini model until the quota is reset.

Pay attention to the word "silently".

There is no pop - up prompt, no change in the model label, and no visual feedback at all.

You still think you're using the flagship model, but on the other side, it has quietly switched to the mini version.

Pro users don't get too excited either.

In the Heavy thinking mode, the highest reasoning level exclusive to Pro users, when the server load is high, it will also be capacity - limited. There is also no early warning.

In other words, a $200 - per - month Pro subscription buys a service that may be "switched secretly" at any time.

This kind of operation of "the label remains the same, but the 'brain' is changed" was caught earlier on the Codex side.

In February this year, an issue appeared on GitHub. A Pro user used the trace command and found that he requested GPT - 5.3 Codex, but the actual returned model was GPT - 5.2.

It's not even 5.2 Codex, but the lower basic version 5.2.

He posted the reproduction command:

• RUST_LOG='codex_api::sse::responses=trace' codex exec --skip - git - repo - check - s read - only - m 'gpt - 5.3 - codex' 'hi' 2>&1 >/dev/null | rg - o --replace '$1' '"model":"([^"]+)"' | head -n1
• Output: gpt - 5.2 - 2025 - 12 - 11
• Expected: gpt - 5.3 - codex

Multiple Pro users confirmed the same downgrade under the same issue.

Moreover, this downgrade is "sticky", it won't recover on its own, and there is no explanation.

Even on the day when GPT - 5.5 was released in April, some users reported that the speed of the Fast mode was similar to that of the Standard mode, but the billing was still based on the Fast mode.

A simple task took 7 minutes and 49 seconds, while it should normally take 5 - 6 minutes.

OpenAI admitted it, and then nothing happened

On May 15th, a record appeared on OpenAI's status page.

GPT5.5 Performance Degradation. We are investigating the issue of performance degradation of GPT - 5.5 reported by some users.

On May 17th, the status was updated to "resolved".

However, judging from the timeline of the forum posts, the complaints about the "dumbing down" from May 24th to 26th were more intense than those on May 15th.

Either the "resolved" problem came back, or it was never really solved.

Every upgrade is a "dumbing - down controversy"

Although all companies may face complaints about "the model becoming stupid", OpenAI has not missed any update from GPT - 5 to GPT - 5.5.

Every time OpenAI says it's investigating, and every time it says it's resolved, but the problem continues in the next version.

In August 2025, GPT - 5 was launched for the first time. The title of a hot post on Reddit was directly "GPT - 5 is terrible". Users complained about short responses, more rejections, and less personality.

OpenAI was forced to urgently restore the GPT - 4o option. Altman personally admitted on Reddit AMA that it was "bumpier than we expected".

In December 2025, GPT - 5.2. The translation quality regressed, it fabricated non - existent APIs, and refused to execute style instructions that 5.1 could easily handle.

In February 2026, GPT - 5.3 - Codex. Pro users were silently downgraded to 5.2, which was confirmed by the trace command.

In March 2026, GPT - 5.4. A post titled "GPT - 5.4 has significantly degraded in Codex" appeared on the OpenAI community forum, and all the replies from netizens confirmed it.

At the beginning of May 2026, GPT - 5.5 Instant was launched. The reply length was shortened by 30%, and emojis almost disappeared. Netizens summarized that the accuracy improved, but the "temperature" was gone.

In the late May 2026, that is, now. Complaints about the "dumbing down" of the Thinking mode broke out again.

Lisan al Gaib revealed that since he led the battle for ChatGPT Plus quotas when GPT - 5 was launched, "he receives such private messages every week".

The latest one was someone asking him to help get back the xhigh/heavy thinking mode.

The day with the strongest performance is the release day

chatgptdisaster.com compiled 1087 verified user complaints. One type of scenario that was repeatedly mentioned is called "routing layer failure". The UI shows GPT - 5.5 Pro, but the output is of a completely different level.

Users described a reproducible pattern. After a long conversation, the model starts to "completely ignore what you say", but the top - grade label is still on the model selector.