Due to caching improvements, we were able to cut the costs in half for these models. Our hope is that the promotional period gave you the opportunity to try the new model without worrying about cost.
* Depending on your plan, you may be subject to rate limiting after heavy use of these models.
Just looking for clarity here, as I am on a plan and have not used all of its offering up yet.
ATM it looks like all of the GPT-5 models available in Cursor are “thinking” (or “reasoning”) type models. Do these kinds of models still require two requests, one for the thinking and one for the actual work?
In the future, if/when non-thinking variants of the GPT-5 models are added. How will that affect “pricing”? I think this is still a confusing area of pricing here.
The “fast” models, since they give your requests priority, cost double. But, since these are thinking models, does that mean “double double”? So two requests for the thinking, then double that for fast?
Can you explain something important what happens if someone is billed not per request? Are those models free for them or not?
It doesn’t sound logical that they would pay for something that someone on a per-request plan doesn’t pay for, but it hasn’t been made clear here.
@beru Maybe they still had to update that, I just used gpt-5-high-fast, with 4x request in the interface description but it only consumes 2 fast request out of my 500 this month.
@yakovw In my case, no it’s not. Because I’m in old pricing. All fast gpt-5 models (low, middle, high) all cost the same (2 fast credits) so I’ll just use the highest of them all, same with the non-fast, I always use high. But yeah, it will be costing Cursor more, so thanks for that I guess.
As for the quality, I had better results in my stack during the free testing week for the high-fast variation compared to the other variants.
Thanks for the update — fascinating to know.
I wonder what Cursor’s considerations are.
This is very interesting.
You’re not the first to claim that it’s actually a higher-quality model behind the scenes, even though OpenAI clearly says it’s the same model.
Very interesting.
They are already considering that months ago, hence they are sunsetting the old legacy pricing model by Sept 15. The per-request-basis is costing Cursor more. I’ll be just enjoying the remaining days of my account.
You’re right, it is the same model, only the thinking capacity differs (low, mid, high). Based on my experience, it is not so good in UI but I am in heavy refactoring of my C# project right now and I find it very good in digging deeper into my codebase. Well, I do pair it with sonnet 4. With gpt-5 for investigation and sonnet 4 for implementation, I had managed to finished refactoring in two weeks which I couldn’t make for a whole month. But this owes to the fact that gpt 5 was free for a week so I use it intensively. Now that it is no longer free, I only tap it on complex situation that needs deeper analysis.
I’m not so sure about this, but I think I’ve read somewhere (maybe reddit) that it is not free in the new pricing scheme, but it cost so little it will help extend your limits per month.
Auto is not a model. It chooses models according to availability and load. So with Auto, you can get any model, some of which may do better for UI, some not. You could potentially end up using a terrible model for UI, if you are using Auto, because Auto is not a model but effectively an arbitrary model selector.
I’ve explicitly used numerous models for UI/UX with Next.js. Sonnet was the best, however I now believe that GPT-5 is even better.
