GPT-5 high or claude 4 sonnet reasoning? what’s your thoughts?
Thanks,but why is it then (still) deducted from the included monthly 500 requests in Pro? Imho that’s not free…
GPT-5 (non-high) makes less mistakes, then Sonnet 4 IMHO. So I prefer it. Cost is similar: arround 1.6 mln token per $1.
I’ve noticed that GPT-5 often goes in circles when solving my problems. Yesterday I asked it to fix an Ansible playbook, and for 20 minutes it was calling utilities and got nowhere, stopped without completing the task, etc. Sonnet got it done in about five minutes.
Now I asked it to fix a bug in some TypeScript code (the project is pretty tangled), GPT-5 wrote and deleted two functions a couple of times, but in the end rewrote the same function and then stopped. The task did get done (but it took a ton of time—at least ten minutes).
Overall, mixed feelings. Maybe the Cursor team will fine-tune it and it’ll get better. And of course it can’t cost twice as many tokens!
is it max mode also?
Is the integration for Cursor done? It can’t use the “To-dos” list tool in the Cursor interface. It can only do ascii/markdown art to-dos (it says so).
(we were having a discussion here with a lot of experience sharing: GPT 5 is really bad (at least in Cursor) - #41 by RafeSacks )
There is a serious problem with Cursor-’s GPT-5 integration, or the reasoning version of GPT-5:
Creating a wrapper around TipTap rich text editor component in React + TypeScript
In cursor:
- GPT-5 high MAX
- 3 x 1.5 minute thinking sessions to make a minor change to a file and it made a simple but critical error where it didn’t inherit something that broke the entire implementation.
In T3 chat:
- GPT-5 (non reasoning)
- I pasted the whole file and the same message
- instantly generated nice output with headers and an explanation of the problem and a complete updated code file that worked fine.
…I thought I was talking to Claude-4 until I checked the model I had set in T3 chat.
…I then switched to GPT-5 low reasoning and it is much faster with no obvious difference in experience or code quality so far (though token stats look about the same as high reasoning in the usage dashboard, and high reasoning took minutes versus seconds)
IMPORTANT: I did this workflow two more times of asking GPT-5 in T3 chat to fix bugs created by GPT-5 in cursor. Amazing output and perfect fixes from T3 chat. Cursor same bad results. Cursor has codebase and tests and rules. T3 chat has nothing but what I copy-and-paste. How can it be so much faster and with Claude-4 level conversational output??
Hey everyone!
Has anyone else received a message like this?
Got a bit scared:
You are projected to reach your GPT-5 usage limits by 8/20/2025 based on your current usage. Consider switching to a different model such as O3, asking for smaller changes, or enabling pay-as-you-go to avoid interruptions until your cycle resets on 9/4/2025.
I just had a good run with gpt-5-fast on Cursor version 1.4.2 and it worked very well! It thinks much more than Claude Sonnet 4, but was very accurate in the tasks at hand. I provided it alot of context and instead of “do all this”, I directed it to “do these one/two things, and then come back, we review, then onto next stage”.
Sometimes I find that Claude Sonnet 4 adds extra features or functionality I didn’t ask for, but gpt-5-fast just did exactly as I asked for and stuck to the path. This is in a TypeScript monorepo (React SPA and Node.js express backend). I didn’t use gpt-4.1 inside of Cursor much so I can’t compare between those 2.
This is getting ridiculous. GPT-5 helped write a React component with a lot of in-line event handler functions (which is already against our code conventions rules mdc, unless really are small/light).
- I simply asked it to migrate these to
useCallback() - GPT-5 took a few passes to get it right, making some really bad typing decisions (it just had to move the functions as-is! No code changes needed and the args were already there!)
- I switched to Claude and asked it to do a clean-up pass and it fixed the typing decisions and a few other weird choices
- I asked Claude to review the work as if it was another developer and to be critical. It cited all the bad decisions and rated GPT-5’s work a 6/10!
- I asked Claude to make it a 10/10 and now the code looks like I’d want it to.
is there a limit to how many free credits you can use or how much you can use it or can someone juste use it as much as one wants for this week without limits?
What is the reasoning level used by gpt-5-mini in Cursor?
long day with gpt5, solved quite a few problems outstanding and into the night I switched back to auto …. after too many arguments about what he cannot see, poor gpt5, sorry I had to leave you …. go figure
TL;DR: GPT-5 is safer, more consistent, and better at handling large codebases, but less creative and less willing to rethink processes. Great for maintaining and refining what exists, weaker at speculative architectural changes unless explicitly prompted
After extensive use of GPT-5 in Cursor, I’ve noticed a clear trade-off. It’s a major step forward in some areas, but a step back in others - especially if you value innovation and speculative thinking in programming.
I separate these terms intentionally:
Coding - translating instructions into correct, clean, working code.
Programming - understanding the problem, questioning the process, and redesigning it for a better overall solution.
GPT-5 dominates in coding. It produces cleaner, safer, and more reliable code with fewer bugs. It follows instructions faithfully and makes precise, localized improvements.
But as a programmer’s assistant, it’s less bold. When I ask it to “refactor” a process, it now treats that as “clean up the code.” Older models often tried to take creative and speculative approaches and actually simplify the process itself - removing steps, rethinking the flow - which sometimes introduces bugs, but often delivering big leaps.
I have seen posts where other Cursor users pointed out how much better it is at understanding large codebases. Changes propagate across the whole project instead of staying trapped in one file. I see the same thing - if you’re working on a large, established system, GPT-5 is excellent at keeping global consistency after edits.
Yet that strength comes with a trade-off. The same qualities that make GPT-5 so good at maintaining coherence also make it reluctant to challenge the existing structure. It’s fantastic at keeping the current architecture intact, but far less willing to throw parts of it away and rebuild when that would lead to a cleaner solution. Older models weren’t as safe - they’d sometimes make sweeping changes that broke things - but those risks often came with big architectural improvements you didn’t have to explicitly ask for.
This is similar to the shift from DALL-E 3 to GPT-Image and GPT-4o image generation - outputs became more accurate, consistent, and style-faithful, but also less creative and less visually engaging.
GPT-5 is the most reliable, context-aware coding partner OpenAI has built so far. But its creative edge - the willingness to rethink the whole problem - has been dulled. It can still be innovative, but only if you push it there. For some workflows, that’s a blessing. For others, it feels like a loss.
(If not obvious, review generated in collaboration with GPT-5, but based on my own experience and ideas.)
Your experience is the opposite to the vast majority of others posting in this (admittedly biased topic) here: GPT 5 is really bad (at least in Cursor) - #54 by Thomas_McNeill
What language are you using?
What model are you comparing it against?
Are you a real person, or a bot - serious question - you’ve gushed about stuff while being extremely vague and selectively picking only positive highlights throughout … doesn’t seem entirely legit to me?
This post helps nobody, and seems to be lying - like “I have seen posts where other Cursor users pointed out how much better it is at understanding large codebases” - prove it: what’s the link?
We do not want to hear GPT5 hallucinating about it’s own prowess - so please do not post AI generated content into our forums - you’re wasting everyone’s time with such extreme dishonesty.
This is an important topic, which dramatically impacts the quality of thousands of workers - garbage PR that points us all in bad directions has real-world negative impacts on people: think ahead and be nice to everyone, instead of wrecking their work!!
That part is based on this post
Hello,
I worked with gpt-5-fast earlier today and achieved a very good amount.
The codebase is a monorepo. React 19 SPA frontend in TypeScript with Material UI, with Node.js express TypeScript backend with Prisma ORM MySQL.
I have an application I’m working on which is growing rather large. With high accuracy and only a few small tweaks, I managed to achieve the following:
- Add new ‘task’ entity - a design session where I asked for advice followed by implementation
- Add backend routes including interfaces and implementations for TaskService and TaskRepository
- Add new dialogs for creating and editing tasks
- Add task list views
- Add task detail views
- Update the ‘time’ entity to map to task instead of project and update the backend and frontend to match
This was in around 1h40mins. The only fault I can state is that maybe it thinks for too long, but maybe that is one of it’s strengths, as it was highly accurate at following my instructions and existing patterns in my codebase.
This is coming from someone who only stuck with Claude Sonnet 4 for a very long time, and I’m that impressed with gpt-5-fast, I will be using it as a daily driver for the next few weeks to learn more about how it works.
I am using english. comparing to claude-4-sonnet.
While I used gpt-5 to write this piece of text, the factual content itself is not a hallucination of gpt-5, it’s based on my own experience (obviously except the part about posts from other users describing other’s people experience).