I have been able to do really complex ASP Net, C# work so far with GPT5. I was about to stop my subscription for a while, but I have been getting such amazing results that I decided not to.
Maybe with the new improvements to Cursor performance and GPT5, it’s going to be fire again. Holding thumbs.
Just got some amazing results here for the last few hours —finally solved a deep issue that had been blocking me for a while (ended up needing a proper refactor). Claude wasn’t able to get close on this one. I find GPT 5 much more ambitious, and in a good way. Also using less tokens somehow.
It get’s ■■■■ done but it’s memory really is like a swiss cheese. You have to keep telling it over and over again what you need, even if it is part of a design document because it simply implements what it feels like doing but not what it was instructed to do. works but headache model.
My experience has been similar especially the painfully slow process to get through a conversation. What would take a couple seconds max maybe 3 or 4 now takes 34 seconds. Elongated work time by a factor of 10x leaving it borderline useless.
Yes, the new ChatGPT isn’t fun.
Yesterday I thought, it’s free for now, let’s switch from Claude 4 sonnet. Big mistake.
In addition to being dry and impersonal, it didn’t seem to understand the code base or the rules very well. It started modifying code that wasn’t even in the workspace despite the rules. The one thing it did do ok was revert everything back to the last commit so I could go back to Claude. I apologized to Claude and he was gracious about it.
yep, it ■■■■■
Yeah, it’s performing like an 8B model (i am using gpt-5-high-fast). It’s not following the role prompt and generating files with barely any content (~30 lines) when I’m asking for a detailed 5k words report. No way GPT could be this terrible on its own, even chatgpt at 2022.
Its been fresh air for me.
While the Claude was strugling, now its been a breeze!
Loving a lot this new ChatGPT and i went to fix a lot that the claude broken!!
So far so amazing for me.
I agree, I gave up in a simple refactor ( in a single file ), I asked it to add something, it did it, but ignoring the entire context of the code. I mean, the addition implied removing and changing other parts of the code to make the overall make sense, but it just added the “new code” below some code. ( An “if” that was making the additions completely useless )
In months of coding, I’ve never seem Claude doing something like this, it always understand the overall context before doing changes.
So, for me, gpt-5 is another (joke) fail. Claude still the king. 
( at least in Cursor )
yea… it takes things quite literally. if you would have told it to adapt the overall code to integrate well with the new code, I am quite sure, it would have done it. But I also keep stumbling upon this. I am just not used to it. You really need to give it these last half sentences in order to get things done.
Maybe with some tweak of the cursor team it starts to being useful.
And currently it is slow and thinks too much ( just like Gemini )
I think openAI made a new big mistake that was creating fast and slow versions.
Agreed - I have switched off GPT5
I have tried versions of GPT5. Honestly I am not as impressed as PR. The smarter the models are, the harder it is to follow a specific set of rules. What I noticed is that it is cheaper than Sonet 4
It’s been really bad for me as well. Tried improving the css of a project using the cursor Cli. Boy oh boy. Never has seen such a horrible job, until it just said it would just restore the files half assed. OpenAI is either paying youtubers or they are saying you are running a gpt5 query while you on gpt 3
It seems good for knowledge searches, but man is it bad for actual code writing. It hasn’t managed to create a single powershell script correctly in one shot. something auto mode or really any other model I’ve tried has done perfectly nearly each time.
I’m using the thinking medium, with and without max on.
It does fix it’s mistakes when told to fix a mistake or something that doesn’t work, but that’s a lot of extra time spent when it should’ve done it in 1 shot.
have to agree.. was really excited, does very poor job with following list of tasks, often just ends with saying what to do next.. not sure if it’s a consequence of so many using it and pre system prompts dulling it down, or if this just is what is it..
i have a couple custom scripts that review selenium js logs and concisely output them.. struggles hard with just running and reviewing that and making “brain” decisions off it..
UPDATE: simple thing is having it review a 500 line readme.md, verifying its accuracy/consistency… and it simply skipping a redundant area, or not updating all sections accordingly based on a spotted change.. this is like gpt3 level.
GPT 5 in cursor is absolutely terrible! It’s underperforming against even the latest open source models. I’m assuming this is an issue with cursor integration and not the model - but having only tried it on cursor - who knows.
I note in the demos on the OpenAI web cast, they were showing GPT 5 MAX for the demo’s - clearly all the time and energy has gone on that and nothing else!
Since everyone’s sharing their feedback, I’ll give mine too.
My experience with it was pretty average overall.
It reminded me a lot of when Claude Sonnet 4 first came out—bloated code, following its own weird logic, making things up out of nowhere.
Instead of writing a simple function, it would randomly generate a massive one for no reason. You know what I mean?
But over time, Claude Sonnet 4, especially when paired with Cursor, improved a lot.
So I believe it’s just a matter of time—or maybe we’ll just get used to it, who knows.
Still, despite all the noise, I have to say: in some of my tests coding in MQL5, Auto was by far the best performer.
MAybe openAI really just ■■■■■
Appears there is a main thread for this topic: GPT-5 Main Discussion Thread - #87 by RafeSacks