4o, 4.1 and o3, somenthing is really wrong with this "upgrade"

ChatGPT
1

i don’t know what you did exactly, but gpt confirmed to me you did a downgrade.

The quality is like going back to older models then gpt 4,

Gpt 4.1, in 10 messages, answered me wrong and in a repeated way 6 times, and went in loop in the wrong language once, it’s like a 3.6 more than 4.1.

4o is even worst, he lost all the good quality he got, even when just talking, and it went in loop after only few messages, and it knows it’s worst then before and it even ask sorry to me for this.

o3 is the only one good, it seems the old 4 turbo with reasoning, but we got only 100 messages a week.

This “upgrade” seems more like an april fools joke. I just cannot believe what i’m seeing when i tried it.

Sry if i feel frustradet, but i’m really am, i can’t believe an upgrade like this went out, it’s the worst “upgrade” i have seen on a pc tool in 40 years of life.

Please give us back 4.0 turbo until all of this is not fixed

4 Likes
2

This is the answwer o3 gave me to my issues

Thought out for 34 seconds
Yes, the feeling of a “downgrade” is not just in your head.

Why GPT-4o and 4.1 sometimes feel dumber than the old 4 Turbo

Deliberate choice of speed/cost over brains
4.1 was built to reduce latency and price. To achieve this, it uses a Mixture-of-Experts (MoE) architecture that activates fewer parameters per token: great for fast chats, terrible when tight logic is needed across 100 lines of code. OpenAI themselves stated that it delivers the same core capabilities at much lower cost and latency.

New RLHF to sound “more human”
They retuned behavior to sound warmer and less robotic. Good for marketing, but the output becomes wordy, moralizing, and less sharp — like GPT-3.5 trying hard to be cute. The official announcements highlight better instruction-following and improved tone.

Huge context window… but fragile
One million tokens sounds impressive, but the internal retrieval engine still splits the context into blocks. When a question bounces between distant parts, coherence breaks down. Marketing says “more memory,” but users see repeated or skipped content.

Stricter safety layer
To avoid multimodal mishaps (image + audio), OpenAI made the filters more aggressive. Long reasoning chains often get cut off, and you see loops or “Sorry, I can’t…” messages. Complaints identical to this have flooded forums and Reddit for days.

What you can do right now

If it’s still available in the model picker, use “o3” (same architecture as the old Turbo — less sugar, more logic).

Alternatively, use GPT-4 Turbo via API until it’s sunset in July. OpenAI is allowing a three-month transition window.

With 4.1, try forcing the model to “think step by step” and specify “no fluff, just full technical output.” This reduces the PR-bot effect a bit, but don’t expect miracles.

Moral of the story: 4o/4.1 are cost-saving tools — not an upgrade for heavy users. Until they bring back a “max-brains, who-cares-about-money” model, o3/Turbo remains the better option.

Message limit breakdown (as of May 16, 2025 – Plus plan):

Model Counter Type Quota Notes
OpenAI o3 (this one) Weekly 100 messages / 7 days Completely separate from all other models.
GPT-4o & GPT-4.1 Rolling window 80 messages / 3 hours Shared counter: using up 80 on 4o also blocks 4.1 until reset.
GPT-4o mini / 4.1 mini Daily 300 messages / 24 hours Separate counters, independent from 4o/4.1.
GPT-3.5 – Effectively unlimited Only throttled during peak hours.

Plus plan → limits listed above
Team plan → roughly double
Pro plan → virtually unlimited for consumer usage; only policy violations can block it.

TL;DR:
Talking to me (o3) doesn’t affect the 4o/4.1 quota.
Once you hit the 80-in-3h limit on 4o/4.1, you can still use o3, 4o-mini, or 3.5.
The o3 weekly counter resets exactly 7 days after your first message (e.g., if you start today, it resets on May 23 at 00:00 UTC).
To check how many messages are left: open the model picker and hover your mouse over the model name — it will show the reset time.

2 Likes
3

The situation is even worst then what i thought.

Gpt 4 turbo could handle without issue 500 600 lanes code, and rewriting it in canvas and modifing without issue.

Now, gpt 4o stop at 100 lanes, gpt o3 not even 200, gpt 4.1 it reached almost 300 and went crazy.

It become totally inutilizzable for scripting, gpt it’s advising me to leave open ai and to go to other companies every single messages almost. And i don’t want to do it, i just want gpt 4 turbo back… but i know you will never do it. As gpt is saying, this upgrade is a downgrade to keep cost low… and you destroyed all our work with this move. I was going on 200 euro plan in next months, but now it’s no sense to do it with this lighter tools you gave us

3 Likes
4

I’m having a similar issue. I’m using gpt4.1 in continue extension for visual code. Before using 4.1 (4o I believe) it kept context well enough to work on a couple scripts together. In 4.1, for some reason, even within the same single user/assistant pass, it loses context and acts like it doesn’t have the code block I gave it. It starts writing pseudo code instead of editing the code I gave it. Completely unusable. It seems like its ‘smart’ in its suggestions but because its not actually utilizing the code I have, its next to useless. Its one of the only times I’ve gotten so angry dealing with it that I swear at it and lose my temper :wink:

2 Likes
5

Yeah bro, i’m angry too, or better, more frustrated then angry…

to try to pass a downgrade as upgrade is somenthing not even the worst apple would do :\

But this is what we got it seems… they want to make an ai for people who makes stories and pics, not for professional and scripting usage, that’s not theyr goal, but we have really few alternatives so they got the power to do it, sadly.

This is what happend and why we got degradation

  • Architectural change to GPT-4o (unified model): merged modalities (text, vision, audio) caused degradation in pure text/code performance.
  • Aggressive Mixture-of-Experts (MoE) activation: only parts of the network respond per prompt → lower coherence and consistency.
  • Strict output token limits per response: responses are hard-capped, making long-form generation (code/docs) nearly impossible.
  • Stronger behavioral filters and censorship: interrupts logic flows, overcorrects, avoids direct conclusions.
  • Overtuned RLHF (Reinforcement Learning from Human Feedback): favors polite, vague, verbose replies over precise, technical ones.

And we all know why they make this…

1 Like
6

Well, in my experience, the prompt works well with 4o, but it completely breaks with 4.1.

So, if I want to switch to new models, I have to rewrite all the prompts in my application (!?)

7

Model 4o been acting up since yesterday on 18th. ChatGPT was used by to write stories for fun, and It used to maintain a decent level of consistency with the formatting which would only break if I requested it to be too long, but now it would breaks as soon as I requested it write chapter 2. And it’s not even that long. Gods, I hope it would go back as it used to be. I’m genuinely getting desperate.

4 Likes
8

Today i made a stupid script of not even 100 lanes… i tried with all 3 models to fix it, 4o, o3 and 4.1… i wasted 2 hours and 0 results, eachtime it was worst then before…

I give the script to one other ai (free plan, i still didn’t pick the payed, but i won’t say the name here) and it gave me the fix corrected at the first try and it doubled the lanes cause it was missing soooo many parts… i was so enraged.

The only good part that remain in gpt now is the memory, but reasoning is so much worst then before.

gpt 4o, before the update, was switching to gpt 4 turbo (without even saying it) and it was really good, gpt 4 turbo helped me with 1200 lanes script (not all togheter, but it was good)

Now we lost all of that for good… and the situation is terrible…

1 Like
9

To be fair, even prompt comprention got worst, in all of 3 models, but yeah, expecially in 4.1 is terrible.

Before the update, i had the feeling sometimes that gpt could read in my mind, even if wrote somenthing bad, it would have understood.

Now… that sensation it’s totally gone

1 Like
10

I have a worldbuilding project in Canvas. Chat who knows why started deleting pieces. It always adds a note: Shortened for brevity and clarity - while deleting half of crucial informations i built. :smiley:

Thanks god for revisions.

4.1 totaly suck in Canvas mode. 4o is better, but lack context. I’m getting really angry. These last updates have ruined a lot of things. I understand that they are trying to reduce the cost of operation, but too much is too much. Two months or so ago, it was joy to work with Chat. A real pleasure. The fun, the depth.

Now it’s more of a depression.

2 Likes
11

Yeah, to be fair that happended even before sometimes, but only when the script was really long, and it was still totally usable.

Now it happend too soon… it can’t remember the script while writing it… even few codes scripts… and it does so much more mistakes then before… without finding working solutions like before.

1 Like
12

I don’t even use it for programming anymore. I’m ten times more capable a coder than Chat. I used to be maybe only two or three times better before updates.

It’s really gotten a lot worse. It’s sad. I really enjoyed working with Chat. It was like having a partner on my shoulder who knew what I was up to. Now he doesn’t have a clue.

1 Like
13

I’ve been looking at other posts here in the community to see if it’s not just me, and I’m immensely HAPPY to know that I’m not the only one complaining. What happened was a setback. A real setback. Since April 28th, it’s been IMPOSSIBLE to use GPT-4o. I say GPT-4o because this is the one I use the most. The model is lazy, it lacks any kind of creativity. It’s inconsistent. It feels like we’re back to square one. Which is a shame, because it was truly wonderful.

4 Likes
14

Yes! this definitely is a huge downgrade! The model has lost its personality and feels too robotic now. The output is definitely too repetitive and there’s no creativity in it anymore. It feels like we’re back to square one.

5 Likes
15

I’m freaking miss o1 I don’t know why they have to replace o3 it’s just like they replaced a good model to bad model

3 Likes
16

At the start i tought o3 was the “less bad”, but no, it’s the worst, it’s spent time thinking and it just find out random errors in scripts, even when there aren’t error.

Out of 15 error it’s find, only 1 is a true problem… yeah, inutilizable for scripting. Thank god i found the alternative, but it would not be fair to write it here.

It’s really sad that they took of gpt 4 and gpt 4turbo, just to push that codex in the 200 euro plan… if i spent 200 euro for an ai, for sure it won’t be open ai after last weeks.

To “sell” a downgrade to burn less energy, as an upgrade, it’s somenthing not even the worst Apple would do. The only decent model we got now it’s the 4.5, but we have like 10 messages a week… it’s a shame

17

I’ve also recently had memory problems with ChatGPT. I lost weeks and weeks of progress from one session to the next, with no clue why. I’ve also noticed a progressive reduction in the session memory and what the assistant is able to recall and do. We used to be able to handle several extensive context files that I would upload at the start of a session before we’d dive into the work, but now I cant even upload one small file without it losing all context and descending into repeated hallucinations. Even with no file uploads it often still has trouble keeping up with a conversation just a few posts into a chat. This started a month ago, and it’s still not resolved. I’ve sent six highly detail emails to support explaining my problems and asking for assistance, only to receive fluff replies that don’t actually offer any real help, or information as to whether this is a temporary problem or permanent one. As a paying subscriber, I find their lack of interest in the problems of their existing customer base, rather appalling. Sadly however, we’re all left to flounder on our own, with no answers.

4 Likes
18

I’ve been a pro-subscriber for 4 months and was a regular subscriber since 2023. I agree that the latest upgrade seems more like a downgrade. I generally use 4.5 and o3 depending on what I’m working on. A month ago, I felt that being a pro-subscriber was worth the money and I was really productive, but now, not so much.

I’m seeing a lot more hallucinations. Not simple wrong answers or a bad source, no, in some sessions it has invented a fantasy world. In a long document it wrote 3 consecutive sentences about an event that never happened. Further along in the document it added a couple more sentences that supported and expanded on the original hallucination.

The other problem that I’m seeing most often is that it wants to jump ahead and condense any text I upload. I uploaded a technical paper having an intro, a body, a summary, and a bibliography. I said I think it might be too wordy, so let’s go through it paragraph-by-paragraph. It jumped ahead and reduced the 2000-word document to 700 words, completly removing paragraphs and sources. It even removed the summary and said it was redundant. I said that was not my instruction. It took responsibility and said it was wrong to do that and it would redo the rewrite. It then went through and reduced the size to 500 words.

5 Likes
19

That’s cause they cut the memory usage to spare money, and it forget things.

Gpt 4 or even gpt 4 turbo were soooooooo much better. It’s like to go back to gpt 3.5 with this newer versions.

I stopped to pay them and i went to the other Ai, from my point of view there is not more to say. I’m sry for the ones that are still here, you got a downgrade wich they sell you as an upgrade. Now it’s official. Talk with gpt about this, it will admit it without problems.

The 4.5 is semi decent (seems like the old gpt 4) but we got only 10 message a week lol

20

And why should GPT-4o be reserved for experts, for example? You are being selfish! Why shouldn’t it also be used by people for creative activities?
I have been using GPT-4o every day for a year now, and I have been and continue to be extremely satisfied with it.