[Guide] GPT-5.5 has been exposed for "fake thinking"—replaced secretly with the mini version in just two hours. For a $200 monthly fee, users bought a "Schrödinger’s brain." The Trace command confirms it, and official documentation has openly acknowledged it. Now, users are纷纷 complaining: OpenAI, who are you trying to fool?
ChatGPT has once again been exposed as "dumbing down"!
Over the past couple of days, it blew up first on 𝕏.
Netizen Lisan al Gaib found that after using GPT-5.5 for one or two hours, it suddenly became sluggish—each request received an instant response, but the quality plummeted dramatically.
However, what is displayed on the interface is still "GPT-5.5 Extended Thinking".
In other words, the tag of thinking is still there, but thinking itself has vanished.
$200/month for a "Schrödinger's model"
On the OpenAI developer forum, a complaint post went viral simultaneously.
Agentify.sh states that GPT-5.5 can suddenly lose its ability to follow instructions.
It proudly announced it was "fixed," but the poor code quality triggered widespread rollbacks.
What used to be an easy UI task with 5.5-med is now impossible, even for the simplest changes.
Upgraded to 5.5-high, no effect. Upgraded to xhigh, still no luck.
Moreover, xhigh used to run for several hours, but it has clearly been shortened now.
As soon as the post was published, the comments section exploded.
Someone directly refunded 5.4.
Some users are using XHigh, the highest tier, but say it's clearly underperforming compared to last week—long tasks frequently fail and it completely ignores workflows.
Some users have reported even more extreme issues: “Even a simple query takes a long time to process, and if you interrupt it to correct the direction, it simply ignores you and continues following the previous incorrect plan.”
That's right, everyone is describing the same phenomenon—GPT's brain has been secretly swapped out at some point.
GPT-5.5's current performance is about the same as 5.3, no exaggeration. At first, it was amazing, but now it's completely unrecognizable compared to the original model.
It's not an illusion—OpenAI explicitly states it in black and white.
For verification, Lisan al Gaib conducted a comparative test.
Using the same account, ChatGPT's Extended Thinking produces nothing but garbage, but switching to Codex with xhigh immediately restores normal performance.
In his own words, Codex is "four billion times smarter than this thing."
Developer Andrew Curran came up with a clever idea—directly ask the model, "What is your training data cutoff date?"
Model response, August 2025.
The issue is that the deadline for GPT-5.5 Thinking is December. August is the deadline for the Instant version!
In other words, he selected Thinking, but the system actually ran Instant.
The model label on the interface hasn't changed a single character, but the underlying model has been secretly replaced...
Ironically, OpenAI actually provided concrete evidence for users in its own help documentation.
According to the official OpenAI Help Center, Plus users can send up to 160 GPT-5.5 messages every 3 hours.
After usage is exhausted, the system will silently switch to the mini model until the quota is reset.
Pay attention to the words "silent".
No pop-up notifications, no changes to model labels, and no visual feedback whatsoever.
You still think you're using the flagship model, while the other side has quietly switched to the mini version.
Even Pro users shouldn't get too carried away.
Heavy Thinking Mode, the highest reasoning tier exclusive to Pro users, is also subject to capacity throttling during high server load, with no prior warning.
In other words, a $200/month Pro subscription buys you a service that could at any moment be secretly replaced.
This kind of "same label, different brain" tactic was already exposed earlier on the Codex side.
In February this year, an issue appeared on GitHub where a Pro user, using the trace command, discovered that their request was for GPT-5.3 Codex, but the model actually returned was GPT-5.2.
It's not even Codex 5.2; it's the lower-tier base 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 does not reverse on its own, and no explanation is provided.
Even on the day GPT-5.5 was released in April, some users reported that the Fast mode was as slow as Standard mode, yet they were still billed at the Fast rate.
The simple task took 7 minutes and 49 seconds; it should normally take 5 to 6 minutes.
OpenAI acknowledged it, and then nothing else happened.
On May 15, OpenAI's status page displayed a record.
GPT-5.5 Performance Degradation: We are investigating reports from some users regarding degraded performance of GPT-5.5.
On May 17, the status was updated to "Resolved."
However, according to the timeline of forum posts, the complaints about dumbing down from May 24–26 were more intense than those on May 15.
Problems either came back after being "solved," or they were never truly resolved in the first place.
Each upgrade is a "dumbing-down controversy."
Although everyone encounters complaints about models "getting dumber," OpenAI has never missed a single update from GPT-5 to GPT-5.5.
Every time OpenAI says they're investigating, every time they say it's resolved, then the next version continues.
In August 2025, GPT-5 was launched. The top Reddit post headline read: “GPT-5 is terrible.” Users complained about brief responses, more refusals, and less personality.
OpenAI was forced to urgently restore the GPT-4o option. Altman personally acknowledged on a Reddit AMA that it was "bumpier than we expected."
December 2025, GPT-5.2. Translation quality has declined, fabricated non-existent APIs, and refused to follow style instructions that GPT-5.1 could easily execute.
In February 2026, GPT-5.3-Codex. Pro users were silently downgraded to 5.2, confirmed by the trace command.
In March 2026, GPT-5.4. A post appeared on the OpenAI community forum stating, "GPT-5.4 shows clear degradation in Codex," and all user replies confirmed it.
In early May 2026, GPT-5.5 Instant launched. Response lengths were reduced by 30%, and emojis nearly disappeared. Netizens summarized: precision improved, but warmth was lost.
Late May 2026, which is now. Complaints about Thinking mode's dumbing down have resurfaced.
Lisan al Gaib revealed that since he led the charge for ChatGPT Plus credits upon the release of GPT-5, he has been receiving such private messages every week.
The latest post is someone asking him for help recovering XHIGH/HEAVY THINKING.
The day with the highest splitting volume was the launch day.
chatgptdisaster.com has compiled 1,087 verified user complaints, one of which repeatedly mentions a scenario called "routing layer failure," where the UI displays GPT-5.5 Pro, but the output is entirely different in quality.
The user described a reproducible pattern: after a long conversation, the model begins to "completely ignore what you're saying," yet it still displays the top-tier label on the model selector.
The most absurd footnote is that the mechanism automatically switching Plus users to Mini after exhausting their 160 messages per 3 hours is described in OpenAI’s official documentation as a “feature.”
Why is this happening? According to Lisan al Gaib’s analysis, the answer is two words: save money.
The tightening of mining power and profitability is affecting everyone. Be frugal everywhere and don’t miss any opportunity to save money.
However, in the same week that GPT-5.5 users collectively filed complaints, GPT-5.6 had already appeared in Codex’s backend logs.
Internal code name: iris-alpha, 1.5 million token context, Polymarket gives a probability of over 85% for a June release.
On one side, users on version 5.5 can’t even maintain a basic experience; on the other, version 5.6 is already running real traffic in the background.
This is the 2026 ASI competition.
It’s getting faster to create new models, but harder to let old models complete a conversation smoothly.
The day with the highest splitting performance is always the release day; every day after that is Schrödinger’s GPT.
Reference: https://x.com/scaling01/status/2058643470357590058?s=20
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by Moses.