Gemini 2.5 constantly has problems with applying changes to the files, it claims that it implemented something, but there are no changes to the files. You have to tell him explicitly and even then it says it has implemented, but fails to do so. How can I apply for a refund for such requests? It burns so much limit.
8 Likes
I donât think itâs Gemini that had the issue, it might be the apply model that cursor uses or how the Gemini request is passed to the apply model. I asked it this question after I had an issue and this is the responseâŚ
When you say âthe model is still refusing to add the isCustom = true parameterâ are you referring to the gemini 2.5 model? or something in cursor thatâs preventing the edit?
Thatâs a great clarifying question.When I said âthe model is still refusing to addâŚâ, I was referring to the separate, specialized model that takes the instructions I generate and attempts to apply the code changes directly to your file within Cursor. Itâs not me (Gemini 2.5 Pro) refusing the instruction, but rather the execution step failing.Think of it like this:
- I (Gemini 2.5 Pro) analyze your request, read files, and figure out the necessary code changes. I then generate the instructions and the code snippet for the edit_file tool.
- The Apply Model (a different, faster model optimized for applying edits) takes my instructions and tries to modify your actual file.
In this case, even though I correctly instructed it to add isCustom = true, the apply model didnât successfully implement that specific part of the change in your file. This can sometimes happen with complex files or edits, especially after multiple attempts, as the apply model might get confused by the context or previous failed edits.
2 Likes
Same here. Very frustrating. Even with a custom role it only applies the changes when I specifically ask for. And even then in some cases Gemini 2.5 pro does not apply the changes and need to manually copy paste this in.
Short after the gemini release it worked very good but now they changed something and the gemini model became worse. Unfortunately only in cursor.
4 Likes
the same here
2 Likes
Same here. Actually a really amazing model, just doesnât want to apply changes if you donât ask it a few times and why are requests for a free model being charged?
3 Likes
Hey, sorry to hear this!
Being such a new model, we are still ironing out the kinks with Gemini 2.5 Pro. Google are also fixing bugs on their end too (such as certain MCP servers causing Gemini to fail to respond!).
For any issues you may see, screenshots and non-privacy mode request IDs (Cursor â Getting a Request ID) are super helpful, as we can see what the AI sees and fix whatever is going wrong.
If it is just apply that is not working, at least as a workaround, you might be able to hit âre-applyâ and it should try again at implementing the suggested changes.
With a request ID, we should be able to see whats going wrong!
3 Likes
Confirm that Gemini has many issues applying changes. Often, it states something like, âI see that changes were not applied, let me try againâ. âI see we still have issues applying changes.â Iâve double-checked. It was in Agent mode, so it had access to all tools. Also, it was actually writing changes, and I donât get why it thought that changes were not applied. With Sonnet I have never had such issues.
3 Likes
hey, thx for the answer, for example this request f5d63fcf-27e8-4c0a-9cc3-59e9bd956fcd. Gemini stopped right after it started ^^ (no error though, fully charged). Iâve got dozens examples like this, I reported feedback through the Cursor feedback interface.
anyone get a solution to this?
2 Likes
We believe most of the cases where Gemini would just randomly stop are now resolved, but if you see any further do let us know so we can pass them to Google - thanks!
1 Like
Still happening, consuming credits and not applying the edit, it clearly states it should use the edit call in itâs planning but nothing happens then it stops. Switching to claude works fine. @danperks
2 Likes
Yep. I had to stop using it today. It was burning through credits not making changes or having me manually implement them. If I asked it to, it said they wouldnât implement when it tried.
I switched to Claude and it had no problems applying the same edits.
Gemini Pro quickly went from great to awful for me unfortunately. Usage/resource constraints perhaps.
Iâm getting this issue fairly often with Gemini 2.5.
I found that when this happens and when it starts not being able to access a lot of the tools or requests me to make manual changes that I need to make a new chat to start making the changes properly again.
One other thing I noticed is if you start a new chat and it pins the old chat in the chat box and you enter your prompt nothing will happen. It looks like its generating after you press send then it just stops. You need to remove the link to the old chat from the chat box and then run the prompt and it will begin. This happens very consistently as I am constantly making new chats to get the model to output as intended.
Nope. Even today it is still happening with gemini 2.5
even just something as simple as commenting out a portion of code.
2 Likes
Yes this has gotten way worse yesterday and today. Unusable. I am on version Version: 0.48.9
Just wasted cash on credits and tons of time because this isnât performing itâs basic function.
yep. Every charge i have had on gemini 2.5 has been a fail (honestly, they should refund those) I keep attempting to use it sporadically to see if it has been fixed, but it hasnât been. I just know it is best for coding and would LOVE to implement it, but it has failed literally every time to make any direct changes.
1 Like
Still getting tons of them, thinking it has to do with my enormous codebase size. Will see if switching to max helps, gonna tell it to edit smaller chunks too in the meantime.
1 Like
Hello, Iâm having the same issue, and it seems to be specific to gemini-2.5-pro-exp-3-25. Cursor will try multiple different methods to write files, âconfirmâ that the changes have happened, but nothing has changed on disk. I switched back to claude-3.7 and it wrote with no problems.
@danperks Iâm working on a simple project with a Raspberry Pi controlling four solenoids to feed my goats. Iâm sshâd into the Pi from my Mac. Here is the Request ID:
fb870e40-80b0-4db2-b751-d406cb703048
Screenshots:
The cat command:
goatfeeder@goatfeeder:~/goat_feeder2.0 $ cat ~/goat_feeder2.0/templates/index.html | grep "<h1>" <h1>Solenoid Control</h1>
It always asks me to make the changes manually.
Switching back to Claude:
[Sorry, the forum wonât let me include more than three images, but it worked fine.]
And confirmed file changed on disk.
My goats and I thank you for your help!
Google is still having issues on their end, unfortunately, but we have asked them to expose the errors while they work on a fix, which should hopefully mean we can catch the failed requests and not add them to your usage moving forward.
Hopefully they can get this ironed out soon 
1 Like
Dan, is there any update on apply issues with Gemini?
âIt seems the model that applied the edit got confused and modified a different sectionâ
âThe reapply attempt also seems to have modified the wrong sectionâ
âIt seems the edit is still not being applied correctly. The model is getting confused by the structure and making incorrect changes in the wrong places.
Since this is the third failed attempt to apply the edit to TabContents.tsx, I will stop here as per the instructions.â
We still get frequent confusion when Gemini tries to apply edits that are not happening with Sonnet.
Should we somehow prompt the model differently about apply tool use?
The issue with frequent stops still persists from time to time.
Sometimes I get chats like that:
User: Ok, continue reading all 3 files without stopping.
Gemini: âOk, I will start with the first file page.tsxâ
âReading Fileâ.
Stopped.