Gemini 2.5 pro performing poorly?

Hi folks,

it seems i’m not the only one who thinks that 2.5 pro performs poorly? I have not yet tried through other interfaces but the one in cursor is extremely frustrating last time i tried it. The feeling i had was that it had no info on my codebase, and had to make generic assumptions about what it needed to do.

i would give it a task, and it would make mock fns and mock data, when it could have written real fns and used real data. it had so many print statements and exception handling that i was not able to scroll through all the text to read the actual code.

7 Likes

2.5 pro is crazy without whatever limiters cursor has on it. Use it through cline or roo code, and you’ll see what I mean.

It’s a shame, really, because the integration and UI of Cursor agent is so much nicer, but I’d rather have a ■■■■■■ interface and a ripper model than the other way around.

1 Like

it’s just not making any edits, it’s writing the full code of pages without actually using it. it also just stuck randomaly like “I’ll apply these changes to…” and then nothing, conversation ended..

6 Likes

I’m experiencing the same thing, but it seems to be a rather new issue. Gemini Pro Max has been stellar for me until today. I just updated from 0.48.x to 0.48.7 @deanrie

1 Like

Same here, but the prompts are credited… :-1:

3 Likes

I constantly get comments like this “It seems the edit applied incorrectly again”. It’s annoying because it seems to understand issues better than Claude, but borks on making the changes 50% of the time.

Could you please Provide me the source of Windsurf is giving you a 1m context window.
Roo and Cline are you using with your own key. You pay for what you do.
Right now you have a limited rate for free. But thats not the case for Cursor etc. thats why it needs to be charged.

3 Likes

Yeah I got this last night, like heres the problem let me work on it…

Nothing

3 Likes

if you’re using gemini pro max and cline
which one taking much cost?
i guest cline is better

1 Like

I may or may not have gone a little wild on this, but hopefully captures the sentiment you’re all likely seeing as well.

2 Likes

tbh i have no clue. I wouldnot be suprised if you work on huge codebase it could be much more expensive in cline . its not that cheap for more than 200k context

It was PERFECT in the very beginning then it got severely nerfed.

1 Like

Are you all on Mac?

I am, and I seem to be running into some type of limiting Cursor is putting on Gemini 2.5 pro despite having a paid plan via Cursor. Whenever I try to use the model, I get a message stating:

We’re having trouble connecting to the model provider. This might be temporary - please try again in a moment.

I can use any other model just fine, and confirmed I have not yet reached my monthly request limit on my account. Its bugging the ■■■■ out of me lol

I use cursor on both Windows and Mac, its probably just a coincidence, but on Windows I seem to have fewer issues with Gemini Pro 2.5.

Anyone figure out how to fix this yet?

I’m on mac, I have cursor free and using gemini 2.5 pro exp and gemini 2.5 pro exp 03-25. For the past week have been having no issues: I have my google ai studio key in settings, have only gemini models activated. You can only use this in "ask’ mode, not “agent” mode with free cursor.
However, I have run into today “user api limit reached,” and I can verify on my google account that I’m no where close to the rate limits. I have had mixed luck waiting for a bit and/or switching gemini models.

I got some waste request with gemini pro 2.5
my gemini 2.5 never applying changes even in new agent chat

is there any bug in latest version? idk
but so far only Claude that perform better in my end

2 Likes

I have to agree, Gemini 2.5 feels better than Claude in the speed, reasoning and overall performance, but it seems handled worse by cursor, compared to Claude. Claude is going above and beyond using tools and getting the information it needs. Gemini seems to simply vibe code its way into solution, botch file edits, etc. Sonnet 3.7 did this at the beginning, so perhaps it’s just tuning the interface to the model? It really needs some work… looking forward to its improvement, it’s really good!