hi @ewzxyh @liquefy those are options offered by OpenAI on API.
hi @MikhailTal @pray4her you have not lost your usage. in the launch week it is not charged but listed for transparency so you can see how it performs and can track usage.
@yakovw all usage of GPT-5 is free for the launch week!
Looks like one of the admins accidentally deleted my last post when I asked which to use. So I will try again. Can you explain the differences between all these that appear in my model list:
gpt-5
gpt-5-fast (thinking)
gpt-5-nano
gpt-5-mini
gpt-5-turbo
gpt-5-high (thinking)
gpt-5-low (thinking)
gpt-5-high-fast (thinking)
gpt-5-low-fast (thinking)
hi @futuretech the models are as offered by OpenAI, you may look at their documentation for the model.
@charles we offer models provided by OpenAI on API. see above.
Why will gpt-5 cost 2 requests per call once the free trial is over? The actual official API pricing is the same as o3 and Gemini Pro 2.5 but you’re intending to charge the same for it as Sonnet 4 with Thinking?
@tommy5dollar we will announce the details before the launch offer ends.
Since yesterday, I’ve had a noticeably better experience using GPT-5 in Cursor, in contrast to much of the feedback I’ve seen from other users.
In particular, when working with very large codebases, GPT-5 handles the work far better than any other model. I believe a big part of this comes down to GPT-5’s excellent tool usage. Together with Cursor, it genuinely feels like the two are working as a team.
It reliably finds the right places in the code, makes targeted edits, and does so with a solid understanding of the overall structure. As a result, changes aren’t just localized, they propagate correctly through process chains, ensuring that every related step is considered and updated.
I’m not sure whether this improvement is due to changes in Cursor’s internal context-finding and selection logic around the GPT-5 release, but the difference feels significant. Edits are more precise, coverage is more complete, and the overall workflow is smoother.
Another major improvement I’ve noticed is that GPT-5 handles weaker or less precise prompts much better than before. It seems to internally refine the request, identify the true purpose behind it, and deliver the best possible result, even when the input is imperfect.
Overall, great job to the Cursor and OpenAI teams, for me, this really takes assisted coding to a whole new level
aside from the pesky usage limit 
One difference I’m really happy about compared to O3—it’s a big deal—is that he talks through and explains what he’s doing. O3 would first do everything—research, check—and only explain at the end. He also explains code, which helps me know when to stop him if he’s wrong.
Have you already reached the limit?
We need the GPT-5 model with minimal effort reasoning. Non reasoning models are the best at agentic coding for specific tasks.
obv. for complex problems or debugging a reasoning model is musthave.
But for exact specific tasks non reasoning models perform much better.
Perplexity has no reasoning on GPT5 and im getting significant better Frontend designing results.
This was the same with Sonnet 3.7 and Sonnet 3.7 Thinking and 4.0 and 4.0 Thinking.
i always got better UI by the nonthinking. This has something to do with the token consumption
also openAI says in official docs minimal reasoning can perform better at coding!
Screenshot 2025-08-07 2351011251×1209 189 KB
@condor Hey, you closed my Thread with the reason its already available. This is not true. Still only reasoning models are available.
I got a notification, but only once. I don’t know how much longer I’ll be able to use it.
Yeah, I faced this issue too.
You don’t have much to lose — use as much as you can.
בתאריך יום ו׳, 8 באוג׳ 2025 ב-13:45 מאת Elmo via Cursor - Community Forum <[email protected]>:
But how much did you use?
For what amount?
בתאריך יום ו׳, 8 באוג׳ 2025 ב-13:45 מאת Elmo via Cursor - Community Forum <[email protected]>:
I want to put my feedback on GPT-5 here what I’ve seen so far and I might add more from time to time.
The different models:
The model is very good, it performs at least as well as, and often much better than O3.
It’s good for:
Building websites, writing code, and debugging really excellent.
Limitations:
If you want to use the model for C# projects or similar things not React and TS, but less common code that’s probably underrepresented in the training data you should use only the strongest model right from the start.
Cons where is the model really bad?
Overall it’s good, but one area where it’s weak (same issue O3 had) is producing answers that are pleasant to read. Interestingly, on OpenAI’s website it does this well, but through Cursor it doesn’t, while Anthropic’s models and even the model in AUTO write nice, readable answers, formatting them with headings so they’re easy to scan. GPT-5 is really terrible at this. The answers come out messy and hard to read. Maybe this can be fixed with instructions; I don’t know.
How about Ultra users?
Does this fist launch week of GPT-5 reduce the avalable amount of token-usage?
What about fast high high-fast models? How does it reflect on my so called “unlimited”.
I use a lot, so I usually can’t use big models too much - but how about now with gpt-5 first week? Can I really test it at full? Or I wll be “punished” further with model degradation because I used to much of, e.g., high-fast ?
I tried GPT-5 today extensively and like it much more than Sonnet 4
Really more
Tasks: Architectural decisions, refactoring, TS, Vue, PHP