I don’t know what the ■■■■ is going on, but Sonnet 3.7 is getting dumber and dumber by the day, to the point I spend more time fixing his ■■■■ rather than benefitting from it. On ther other hand, Gemini 2.5 keeps making fun of me, where he constantly refuses to edit the files. I need to ask 3 times in a row to convince it to actually edit the ■■■■ files. It’s frustrating.
Same problem
3 Likes
When I first tried Cursor it was great, since using it and upgrading my account, i’ve had I think 3 or 4 updates and each one has made the AI more useless than the last.
9 Likes
Same, wenn i compare the plain models with the models within cursor it’s a clear difference. Plain Models on claude app and ai studio are amazing good, the same models in cursor are pretty dump. Sometimes im not sure if it’s really using the selected model. Because the output und the understanding of the prompts are so bad. Feels like gpt-3.5 sometimes. 
1 Like
Same here, frustrating.
Not only Sonnet, every model after a couple of days of the release date gets dumber ( & now Gemini-2.5)
3 Likes
Same here! ~ 3-5 days ago I received pretty good responses, now cursor is totally dumb.
What’s going on cursor team?
7 Likes
Same here
The model’s acting dumb because Cursor uses its own language to create prompts (like JSX). They tweaked it a lot to work with the model’s thinking (especially after Sonnet 3.7), making other models dumber compared to using a custom API.
very much so. results have deteriorated after being intially stellar. maybe they did NERF the models
1 Like
Hey, can anyone provide any clear examples of where Sonnet 3.7 or Gemini 2.5 Pro are being dumb? Non-privacy mode request IDs are best here, as we can see what is missing from the Agent’s context and what Cursor should be giving the model that it doesn’t already have.
To clarify, we have not nerfed any of the models, context windows or token allowances inside Cursor.
The context windows for each model have never been higher, and are published here for visibility now:
3 Likes
I’m sorry, Dan, but a lot of people around the world are reporting the same problem. It’s hard to believe that it’s not getting worse when I see so many people with the same idea. Something isn’t right with Cursor. I’ve been using Cursor since the beginning, and some updates were really great, but the recent updates haven’t been as good. Since yesterday, I’ve been trying to solve a problem—almost 24 hours—and despite many requests, even premium ones, it’s not getting fixed. Instead, it’s just creating new issues.
6 Likes
For me, Claude says there is no Claude 3.7 sonnet only Claude 3.5. This is even when I clearly only have Claude 3.7 sonnet activated in my Model Names of my settings. When I ask how to get support, it says this forum is not valid.
1 Like
Hey, models don’t know anything about their version unless it’s specified in the system prompt. The response indicating it’s the Claude 3.5 model is an error that could have been caused if you asked this question in an existing chat where a different model was previously used or triggered by something else. To prove this, I’m attaching a screenshot. As you can see in another screenshot from the Anthropic console.
1 Like
Hello everyone.
I encountered a similar situation yesterday. Gemini 2.5 introduced an error into code that it had largely generated itself. Rectifying this required considerable effort: it took 10 iterative attempts, utilizing both Gemini 2.5 Max and Sonnet 3.7 Max, before a working solution was finally identified on the tenth try.
1 Like
experiencing the same thing with sonnet
Gemini last week was perfect. Now it is useless. What is going on?
It’s gotten really bad overall, no matter the model. Cursor now fails all 3 of the benchmrk tests I use to evaluate if they’ve changed the way the app works. It used to pass all 3 up until about 3 weeks ago.
I’m worried Cursor may be running out of $ or something. People on Twitter are ■■■■■■■ the recent performance, so the sentiment extends well beyond just this forum.
The insanely tiny context window just isn’t cutting it.
Edit: The censored word above is literally the verb form of dog. I guess that’s a swear on here. 
4 Likes
Same experience, for some reason Cursor messes context a lot and responses I’m getting look too far from what was expected.
I can’t see much difference between MAX and regular premium model requests. It feels more like MAX ones are performing even worse than non-MAX ones.