These days, I use Aistudio for my projects and I can achieve what I want in just a few attempts. For the same project, I tried to make a very simple update in Cursor by selecting Gemini 2.5 Pro. Since it normally sees the entire file structure, I expected it to perform better, but the opposite happened — it broke the working part. For the same issue, I added the files to Aistudio and entered the same prompt. I took the outputs from Aistudio and placed them in the necessary spots. When I built it, even though it didn’t work exactly at the level I wanted, it didn’t break any working parts. I didn’t do a code review myself, but the main logic worked.
what is your experience ? Does Cursor reduce the performance of the model?
The test is not very correct, and perhaps the second time was more lucky than the first time.
In any case in Cursor:
- Gemini 2.5 Pro works chaotically and uses tools poorly
- Grok 4 broken from the release
- Kimi K2 maybe broken from the release
Works good only GPT-5, Grok Code Fast, Claude Sonnet Family.
Grok 4 Fast Reasoning lazy or works chaotically.