Using o3 and o4-mini for coding has been extremely disappointing.
When you ask for full code, they say things like “Paste the rest here” or skip entire parts.
It’s lazy behavior — and honestly, it feels like they’re cutting outputs just to save cost.
Output length is drastically shorter than before.
o3-mini-high could generate thousands of lines, but now even a few hundred is too much.
This isn’t an upgrade. It’s a downgrade with a shiny label.
Anyone else experiencing this?
I completly agree it s not even. Suitbale for coding
It’s terrible for coding, this is bad news for me cuz i’ve got exams and coursework coming up very soon, this is the worst timing for them to pull this crap
They know what they’re doing not making o3-mini-high available too? If i’m being honest sometimes it’s hard to not see this as them having an agenda. Because the likes the legacy models like 4.5 and 4 remain usable, yet their clearly best model at reasoning, o3-mini-high, no where to be found..?
Hugely bad! Please restore o1! I might cancel my pro subscription ….
Yep, worst model yet. I simply asked for some analysis of a situation and it to detail some stuff, a task I used o1 for all the time. Gave completely useless information. Both are the worst ones yet.
Well said! o3 is a downgrade from even o3-mini which is quite ironic!
Truly disappointing as I was an avid user of pro-o1 and it really helped me. A project I’d been working on this week I’ve now wasted probably 15 hours due to implementing a lot of code from ChatGPT o3 which I later learned was totally wrong. If ChatGPT needs to save money it would be so much better for everyone to just limit the number of queries than allow for infinite queries all of which are providing trash results. Incredibly frustrating because I blocked off this weekend to do a project only to spend 10 hours implementing garbage code all of which I had to revert.
This is very true. The funny thing is, I bet, if they nerfed it to reduce costs, it actually just creates this huge spiral of rework demand - as people hit it like 10x to get the one result.
Until, they all leave the platform. Forever.
So, it’s actually kind of ironic — the company stood out, led the market, gained a lot of experience, and became unique in this space. It created clients that became dependent on it.
And suddenly, it rolls out a nerf that makes it impossible for specific clients using code to keep working.
Look, it’s not logical. Sure, it makes sense if the goal is to reduce costs, but if that were the case, I strongly believe the company would simply raise prices while still delivering a quality product.
So, I believe this is a NERF on specific accounts that go beyond some quota. Because this level of downgrade makes no sense otherwise.
I’ll buy a new account and run tests on O3 Preview, O4 Mini, and O4 Mini High — checking if they can think for more than 1 minute when needed, and deliver at least 1,000 lines of code without chopping it up, shortening, or splitting it into pieces.
I’ll report back. That way, at least, we’ll know if this is a damn nerf on specific accounts or just a shot in the foot (and I don’t believe a company that’s ALWAYS EVOLVING would do this in an UPGRADE).
That’s really interesting. Wondering if it is that. I’d love to hear what you come back with.
I’m getting tempted to leave, frankly it’s getting less and less productive. Also, support now closes conversations without even sending a response on their support chat, their only means of contact, which frankly doesn’t support at all.
Lol I also just reverted a whole day’s work, damn, those new models are trash, barely usable. On launch day the o4 looked great, but a day or two after? Total unusable crap. I wish they’d bring back o3-mini-high…
I am here to post the same thing, o4 models are the laziest ChatGPT models so far, I doubt this is intentional by OpenAI to save computational power
The thing is — it makes zero logical sense!
ChatGPT itself ran benchmarks comparing earlier versions, and o4-mini / o4-mini-high were proven to be better than previous models.
So we say: “But the results I’m getting are complete GARBAGE.”
It doesn’t give correct code, it restricts output, sends broken code, sends pieces of it, says it’s complete when it’s not, doesn’t understand shit — etc.
So why the hell is this happening?
Option 1: The damn account is nerfed.
Option 2: The model is hitting some internal token limitation, causing it to hallucinate or screw up responses.
But if that’s the case — considering how simple the code I tested was — then no one can use ChatGPT to code anymore.
I don’t think the issue is an actual nerf in the model’s reasoning capability, because it’s been tested and shown to perform better on paper.
You feel me?
But honestly, it just doesn’t make sense — and it wastes a ton of our time, drains energy, and stresses the hell out of us.
I’ve already given up on the changes I had to make in my projects. For real.
I’ve been unable to get ChatGPT to produce any working Python code for the past few days.
Its ability to handle coding tasks has deteriorated dramatically. Previously, using the o3-mini-high model, it took me just a few minutes to create a working mini-application.
However, after the recent ChatGPT update, it seems to have completely lost its coding abilities—I could spend an entire day without getting even a single line of functional code.
Interestingly, out of desperation, I registered a new free account and asked ChatGPT to do the exact task I’d been struggling with all day using my PLUS subscription (o4-mini-high model). Surprisingly, the free version of ChatGPT completed the task in just a couple of minutes! Unfortunately, my usage limits for the free plan were immediately exhausted afterward.
Initially, I thought this issue was unique to my account. Now that I’ve found this discussion thread, I realize it’s affecting many others. If not all PLUS subscribers are affected, but only some accounts, there’s a risk of subscribing again on a new account and ending up in the same unfortunate situation.
I’m looking forward to hearing from our respected ESTUDANTE_PICASSO, who seemed to be planning an experiment with a new account.
Hi everyone! As I mentioned, I bought a brand‑new account…
But the file still gets cut off—same disappointing result.
Even a new account doesn’t fix this nerf issue.
Old account: estudante_picasso@hotmail.com
New account: estudantepicasso@gmail.com (current)
I’ve tweaked every setting I can find, but ChatGPT still won’t deliver long code blocks.
Maybe splitting your project into multiple ~200‑line files instead of one ~800‑line file is the only workaround.
In short, using ChatGPT now means overhauling your entire workflow and forgetting years of habits.
Don’t waste your money on new accounts—it doesn’t help. Now I’ve got to chase down my refund.
From 725 lines, I only receive 345—as if that were the complete code.
That’s it. Save your time and cash: don’t buy another account.
I tried posting this from my other profile, but it needs approval, so I’m sharing it here instead.
The o3 for me does exactly what you said. It abruptly stops printing. It didn’t care for my prompt and have an approximation. But after 10 -20 prompt poking,it gave a correct response. But any number of prompts and it didn’t complete the code.
It’s very, very bad now for coding. Feels like GPT-4 Turbo level, or close.
o3-mini-high was miles ahead of the current o4-mini-high or o3.
These models seem heavily gimped, possibly due to compute constraints? They don’t really think anymore - they feel lazy and shallow. Hellish experience after o1 and o3-mini.
How does that reduce costs, if now I need to enter 10 separate requests to solve a simple task that took only 1-3 prompts with o3-mini-high?
Who’s going to give me back the time wasted debugging the nonsense these newer models output?
And then I have to submit 30 prompts just to get the simplest code working? That’s absurd.
I genuinely can’t believe that the amazing engineers at OpenAI - the same people who delivered the excellent o3-mini and o1 - are behind these current models.
It’s astonishing and mind-boggling how such a downgrade made it into production.
How did this even pass QA?
If it did pass, then… is this level of performance actually intended?
Are there any updates from OpenAI?
Any fixes coming?