Hey everyone,
I’m not sure if it’s just me, but I’d love to hear your thoughts.
I’ve been using GPT-5 heavily on Cursor since Thursday. For logic, huge problem solving, it feels like the best model I’ve used so far—and I’ve been using Cursor for 9 months heavily and straight.
Opus, is really good but the pricing is absurd… so i hate it kkk
Don’t get me wrong, Claude is definitely miles better for design, but I’m seriously impressed with GPT-5’s coding logic. Am I the only one feeling this way?
I give up GPT5 even though it’s free for now after use it 2 hours , I think this mode coding capability is not as good as Claude 3.5 in cursor
While I have had a fair number of problems with GTP-5, and it has tried to run some SCARY commands so far (things that would have lost code changes or other data and content)… So I am quite wary of it.
I switched to the gpt-5-fast model yesterday, which seemed to quell a lot of the rogue behavior that I was seeing with the main gpt-5 model (honestly don’t know why the two seem to behave differently, but they do…at least so far). That gave me a chance to see what GPT-5 could really do.
I am not sure that it is so much coding, as it is…GPT-5 might actually understand software architecture, principles, patterns (well, actually, I guess I would call that DESIGN) better than Claude 4 Sonnet. Sonnet does a wonderful job, and it does seem to act more like a junior to mid level developer that actually has some skill most of the time.
GPT-5 though, when its not doing something rogue and unwanted, does seem to generate better designed code. It also seems to recognize when my own instructions, might lead to less optimally designed code, and it will note such. Sonnet will also do that, but not quite as often as I have noticed gpt-5-fast doing it, and…I do think that gpt-5-fast is a little better at it.
Overall, I don’t think I’m seeing anything revolutionary. Little steps, which is what I’ve come to expect from each new generation of model. Claude 4 vs. Claude 3.5/3.7, was IMO a much bigger jump than Claude 4 to GPT-5. I guess we will see, when Claude 5 arrives, how much of a jump it is. GPT-5 does seem more incremental though.
One thing about GPT-5 that does seem much better, is speed. It is just a lot faster than Claude. Claude seems to really throttle their output rate. GPT-5 screams.
gpt-5-fast has been amazing in my experience these past days since release. It’s the only model I’ve been using. I’ve used it for refactoring existing code and also adding new features (but following very consistent existing code patterns) and it is pleasant to work with. Great work.
Think of it as o4.
It’s got quirks, but has proven crazy good for debugging for me.
It’s really good… was into a coding problem it couldn’t solve, switched back to claude-4-sonnet, and it couldn’t even grasp the issue.
Also, asked a free ChatGPT how to fix an RDP issue: it eventually had me delete a piece of the Windows registry that it said would get recreated upin reboot: it wasn’t, and my server was down… spent a few hours with that free ChatGPT trying to resolve it. Turned GPT5 loose on it (didn’t realize I could ask system issues in Cursor until this time), and it took a few hours, many reboots, hundreds of commands, but it finally got it working again.
What’s it going to cost when it’s no longer free? I’ve been happy so far with claude-4-sonnet (not max nor thinking) and staying within budget.
I do wish “thinking” mode would not delete it’s “thoughts”… I would like to go through them.
The dialog in “thinking” reminds me of the best Star Trek ever (The Cage, the original pilot), where the Telosians were telepathically communicating with each other while watching the human reactions in their virtualized world.
The thoughts are not deleted. You can click on the text that says “Thought for Xs” and it will show the thoughts again 
Im impressed too, I don’t get why some have problems here!
I just wrote that in another thread:
i don’t get what problems you all have, I switched from “Claude sonnet 4 thinking” to GPT-5 and i can just say WOW! please don’t change anything! It works so well, never had such perfect code, it every time knows exactly what my code is about. and never makes mistakes, I mean really never…. and my project has around >2k files with around 300lines code each file…
im really impressed. what a love the most it uses the MCP tools every time correctly, while Claude every time forget about using them… furthermore it doesn’t change code if not necessary, Claude was like everything is solved with changing code… GPT-5 knows when I made a mistake or when a configuration is wrong…
i just started a new complex project and it made it with just 2 errors, which it solved like a charm…
Love it!
It’s been working great for me. I was working with Claude-4-Sonnet before, and it was pretty good, but it would take several iterations to get things exactly right. GPT 5 gets it on almost the first try. Also Claude-4-Sonnet would go into loops trying to debug something, forgetting that I had already tried that. I haven’t seen GPT-5 do that yet, but then again, even on debugging, it usually gets things right the first try, so there’s been no opportunity for loops.
I haven’t tried GPT-5-fast yet, but based on comments here, it sounds like I should try that out.
BTW, someone mentioned “rogue behavior” with GPT-5. I saw quite a lot of that with a lot of the earlier models that I experimented with. Claude-4-Sonnet was better at not doing this, but what really helped was putting a rule in the “Rules & Memories” section of the Cursor Settings.
I’m not sure, but I think this is a new feature (at least I didn’t see it before). If so, Kudos to the Cursor team for paying attention to user feedback. This has helped my development process immensely.
It is just you. Try in CoPilot. Slow, but same mistakes.
The rogue like behavior thing was me. BTW, gpt-5-fast has worked pretty well for me. Its worth a try, I think!
First, FWIW, I have a lot of rules now. Over time, I’ve bene building up a ruleset, and its about 22 or 23 .mdc files now. Every time I encounter real problematic issues, I have Cursor write a rule for me. It does help keep the agents corralled, or at least better than not having them. Still, I had probably 18 rules on Saturday, and man, the models were still running rogue…
I definitely had some VERY ROGUE behavior, well I guess it was both Friday and Saturday. Saturday had other issues with “conversation length too long” eventually preventing me from using the agent at all with Auto (don’t know if that issue is fixed yet, haven’t heard it was.)
But GPT-5 is not the only one that I thought was going rogue a lot on me. I call Gemini 2.5 “The Bulldozer” because IT, loves to go rogue. Now, when using GPT-5, I had some dangerous commands crop up, things that would have cost me data, code, etc. They did not run (thank you allow lists for terminal commands!), but they scared the ■■■■ out of me the first couple of times before I realized they had not actually run. The first one, I thought I’d lost my entire git workspace (the only real at-risk code when using git!) due to the model (I think it was GPT-5) trying to run git commit -- src/ (Almost ALL of my changes were in the src/ directory at the time!) I stopped using the default gpt-5 model after that, and went to sonnet for a bit. I then heard people having good results with gpt-5-fast and tried that (so far, so good… 
)
Gemini, at least so far, hasn’t done anything outright heart-stopping like THAT, but, it is a freakin bulldozer. It just likes to ‘doze over my codebase, most of the time I use it. It seems hyper-opinionated, and WILL do things ITS WAY, or bust. It doesn’t like Claude 4 Sonnet code, and just wants to change it. All the time. So, I stopped using Gemini to code. I am fine creating plans with it, it works well there, but I DO NOT let it change code most of the time. I suspect, that there was a period of time I was on Auto mode on Saturday…it may well have been Gemini that did some of the rogue code changes, but you can’t know for sure anymore with Auto mode (at least, with one of the recent updates, it seemed like the agent would not let model-query type prompts through and would just respond with a canned message…not sure if that’s still in place.)
But yeah, I have experienced too much rogue behavior from models. I generally use C4S for coding, maybe now it will be C4S and GPT5Fast, Gemini for planning (although GPT-5 Fast plans really well, too), GPT4x for simpler questions and basic research. I’ve used Gemini for research a lot, will probably use GPT-5 for research as well. I kind of have a tough time with Auto mode. I would like to use it, but, it just includes ALL models, you can’t exclude any, so you just never know, when Auto-mode requests are gonna bulldoze your codebase. Would be nice to have some exclusions in there, so when you KNOW a model is just NOT suited to coding, you can prevent auto from using them. Then maybe, it might fall back on something a little slower, say Kimi K2, but IMO that would be preferable to having Gemini stomp all over your code and tell you it ■■■■■… 
GPT 5 is actually much better than its predecessors in terms of programming. But you might be surprised at what Claude can do compared to GPT 5.
ChatGPT is incredible when it comes to time. After waiting 204 seconds I tapped the stop, changed to Claude 4 sonnet and identified my problem in about 20 seconds. It then went on to fix it, build, success and pointed out other code areas that had the same potential problem. So yeah, incredibly slow.
I was able to develop whole portal with Claude 4 in two days, it was so good I started to let my guard down and did not review all the code, then I switched to GPT 5, kept my guard down. Now I am fixing my code for 0.75 manday back with Claude 4.
Claude 4 >>>> GPT 5 in my experience. Although I believe GPT 5 will be great “senior advisor” where Claude runs out of steam.
But the beauty of Claude 4 is that it works really well with Cursor and it does not go on big adventures and does not start rewriting large portions of code.
My 2cents
I don’t GPT5 is any close to Claude 4 especially for creative tasks.
I have been using GPT5 as my daily driver for now but I don’t see myself using it when the free access runs out. 4x with GPT5-fast is simply not worth it. and now, that Claude 4 is 1x, the choice is even easier.
GPT5 is hands down the best for coding, period. If you don’t see it, you’re either coming in with bias from the start or just not really getting how to work with it. I’ve been around since 3.5 Sonnet days to Sonnet 4 with Gemini no stop and I honestly don’t know what kind of magic they baked into this model, but it’s an absolute beast. And I’ve never been an OpenAI fan for code at all lol for me openai always been more for social psychology aspect of things.
Claude code to get things done, and pure joy of using. GPT-5 high for tricky stuff.
How can you say that? I tried it in both Codex CLI and vscode, and in the former it is 5-6 times as slow as Claude Code with Sonnet 4 / Opus. In vscode, it is even worse.
It is useless in a fast pace coding scenario. Anyone who has tried the two should see the same.
I don’t care for Claude Code/Codex/VSCode. This is Cursor. Completely different beast. I was a Claude Sonnet 3.5, 3.7, 4 only user, and since GPT-5 appeared, I haven’t switched back.
Here is a thought.
Expectations are crazy now. Think before LLM’s, Cursor, We didn’t have these massive leaps in technology.
Now when a LLM is slightly slower, it is deemed as worse 
Yes, GPT-5 thinks for a while longer, but for me, it’s results have been superior and it follows my prompting accurately to the point. 
I’m with you. Far better than other, but I still reached out to Claude 4 or Grok 4 two times. In overall, GPT-5 is the best for me right now.