thank you, appreciate it. Personally it responds imo like 4O and not a reasoning model. generic advice in Chat etc. I haven’t found it useful yet.
I’ve been trying to use o3 in cursor but honestly I am extremely disappointed overall. I know we now have the high version, and I am seeing that it’s being less lazy and actually writing files since the tweaks from the cursor team. It’s such a weird model, I think it’s basically o1-mini on steroids with the reasoning postraining. It can be very impressive if used on chatgpt.com but in the context of cursor it feels useless, and also much slower than through openrouter/chatgpt (there you get 1600 tokens per second - right now claude seems faster). I experimented using it as an architect and switching over to claude, but as things stand - it’s much better just to use claude. I assume the underlying reason is it’s a tiny 10BN model and just can’t do tool use properly and quickly loses the thread of conversation. It’s best as a one hit wonder on chatgpt.com. I suspect with a stronger base model i.e. o3-full/o3-pro it might be a different story
I think Cursor is still trying to adapt, because I’ve noticed the same thing—it’s not working well. In addition to issues with context awareness, editing, and applying changes, I’ve also noticed several inconsistencies when requesting a response. Sometimes it seems to reason properly, while other times it returns something immediately, but without much coherence. At times, it’s impressive, but other times, it’s actually worse than GPT-4o.
Yesterday, they announced that reasoning was set to high by default, and I noticed it worked better—slower, but better. You could even see part of the reasoning chain. However, now that part no longer appears, and curiously, responses are much faster—almost instant—but with a noticeably lower level of accuracy.
Maybe it’s Cursor’s implementation, or maybe OpenAI initially launched something powerful and then downgraded it due to excessive resource consumption (like what happened with o1 Pro). We don’t really know, but I do see that ChatGPT’s responses seem superior to what I get in Cursor.
Edit: In my case, some of these issues have also affected r1 and Claude, as well as the workflow in general. I feel lately that I’ve been spending more time trying to get Cursor to work or searching forums for bugs and fixes than actually benefiting from using it. I strongly believe we should have stable versions that at least ensure a functional workflow to a certain extent and prevent new updates from breaking it for a while, as happened with DeepSeek.
I second having a stable/unstable branch to opt into
the api cost of o3 mini with its capabilities is unreal for openai but i encountered it has some problems when working with kotlin. Deepseek works pretty well with kotlin.
Cursor team is goated
It’s the same cycle every time a new model comes out. People rave about it, benchmark show how “better” it is. Then reality hit and everyone run back to Sonnet 3.5
I want to know why I can’t use my $200 OpenAI pro account to access O3-mini-high - why am I limited at all?
That’s true, but in this case I see that o3 could be a little superior for certain analyses, beyond the benchmarks, but also its attempt at integration into cursor has broken the work pipeline also for Sonnet or r1, at least in my case.
Well said , it is mostly all hype and no substance. Lol
@Undisclosed Because OpenAI doesn’t allow to use their account that way. They have a API service and a web service. Your account is to use their web service. It’s not a limitation on Cursor side, but on OpenAI side.
any idea how to enable o3-mini-high?
Can you talk more about what you mean when you say Sonnet being smarter in Aider or Cline?
Restart. Settings / models
o3-mini
in Settings
> Model names
is o3-mini-high
(o3-mini
with high
reasoning level).
I just wanted to report that I have been playing with repoprompt https://repoprompt.com/ and o3-mini is a dream now, so this is definitely a cursor-specific issue. What I love about repo prompt is I have much more control and visibility over the prompt which is generated and the grounding files, it even works for o1-pro if you dont mind copy pasting the generated prompt and xml result back in. The workflow is pretty nice too because you can ask followup questions on chatgpt.com or in the built in chat and it will still give you schematised updates - really nice workflow
Given how well it works there, and given how simple their prompt is, I really wonder whether Cursor are at all interested in making o3-mini work?
I’m still seeing o3-mini Composer (agent) miss a change and then talk at length about the change it made…there’s some dissonance between what the composer is doing and what it says it’s doing.
how can i use Claude 3.5 sonnet for free, im a student
is there any plan for students for free especially Claude 3.5 sonnet