Gemini 2.5 Issues - Looping, Bad Performance, Wrong Language, etc

Cursor used to hide Gemini 2.5’s CoT under an expandable. In tasks where thinking mode first created a plan and is following the plan, this is really crucial since the CoT maintains the latest version of the plan in context, without the user (me) having to re-read the whole plan as it moves through the plan.

Since yesterday evening (basically after the 2nd 0.49 update was released), the entire plan is repeated every time agent goes to think and comes back. So during a task where the Agent had to think 14 times, I see 14 versions of the already verbose plan without an option to hide it.

Gemini is already quite verbose compared to Sonnet 3.7, which is a good thing when working in YOLO mode because I can basically keep reading the output and only expand the plan once in a while. This “bug” made it impossible to follow Gemini since the output is now 10s of pages long for a complex task.

I also noticed Gemini 2.5 performance dropped considerably at the same time. I think (or speculate) this has to do with the very verbose CoT that now leaks into the context, decreasing the SNR of the overall execution, hence confusing the model from identifying the right next step. The performance difference between the yesterday and before is so big that I basically cannot use gemini 2.5 when it used to be my default model before.

Thanks for reporting a bug you have found in Cursor!
Please add the following info to help us diagnose your issue:

:white_check_mark: Check the forum to ensure the issue hasn’t been reported already: No

:lady_beetle: Provide a clear description of the bug: see above.
:counterclockwise_arrows_button: Explain how to reproduce the bug (if known):
Just use Gemini 2.5 in thinking mode for a task that will require planning first, then running a terminal command, checking its output, make a decision according the output and iterate over the plan.

:camera: Attach screenshots or recordings (e.g., .jpg, .png, .mp4).
Confidential so I cannot share, but basically, a single prompt causes a 40+ page (in Google docs) output with the same 4 step plan repeated 10+ times.

:laptop: Tell us your operating system and your Cursor version (e.g., Windows, 0.x.x). :
Mac OS X Sequoia 15.3.2.

:prohibited: Tell us if the issue stops you from using Cursor:
It stops me from using Gemini 2.5. Combined with the fact that Sonnet 3.7 thinking usage-based requests started charing 2 requests instead of 1 (*), this is a pretty bad regression IMO.

(*) I reported this to [email protected] recently. I am not complaining about pricing changes - I am just surprised there was no headsup, nor an update on the pricing page.

14 Likes

Thank god someone else is reporting this.

I was concerned it was a project configuration issue on my end.

At exactly the time of day this user reports, Gemini 2.5 Pro performance fell off a cliff.

From one request to the next, it went from the best available model for my workflow to being literally unusable for almost any purpose, and for exactly the reason @kaykut describes above.

This is a real problem, and it is severe.

I hope it’s a simple enough fix for the Cursor team to restore it to how it worked 24 hours ago.

11 Likes

Me as well same it goes into loops

1 Like

same


He’s bad today A single conversation outputs what’s in this image 15 times… a lot of repetition.

v0.49.4 ubuntu20.04

@deanrie Hey, man, let’s see if there’s something wrong. Looks like we’re not alone. I recorded a one-minute video, but it won’t post, it’s only 8MB max.

1 Like

I used Gemini 2.5 to ask, but it’s using the Sonnet 3.5 model (which I used in the previous conversation). WTH?

it also just goes in loops

1 Like

The model’s response doesn’t match the model I requested - Bug Reports - Cursor - Community Forum

Hey all, we’ve not changed anything on our end here, but would love to figure this one out!

If anyone has any non-privacy mode request IDs they can share, that would be super helpful!

d9a46431-709d-4788-896e-14b69b7f3035


Serious language confusion today… Every conversation is multilingual…

Version: 0.49.4
VSCode Version: 1.96.2
Commit: ec408037b24566b11e6132c58bbe6ad27046eb90
Date: 2025-04-22T00:17:03.731Z
Electron: 34.3.4
Chromium: 132.0.6834.210
Node.js: 20.18.3
V8: 13.2.152.41-electron.0
OS: Linux x64 5.4.0-192-generic

2 Likes

So far no problems with other big models only Google and sometimes it loops wirelessly

I recorded a video but the forum can’t post it 1 minute 90MB

yeah many agent calls just end in a doom loop

This wasn’t happening yesterday. Someone must’ve done something with the model selection.

We don’t think anything has changed with us to cause this, but are escalating with the Gemini team to see whats happening here!

2 Likes

I just asked Gemini 2.5 MAX to update 7 files - quite simple changes, the same thing in each file.

It ended up being 8 new lines in each file.

The response from Gemini 2.5 MAX in the Composer however was over 100,000 characters of back and forth (with itself) with Planning, Analysis, Replanning, as well as a Constraint Checklist & Confidence Score which I haven’t seen before.

Is this a Google change, Cursor issue or something else? It doesn’t seem efficient whatever it is and I hope I’m not paying for it.

2 Likes

Same here, I am getting tons of premium tools calls and keep repeating the same plan over and over. I just spent $40 on something that should have been simple.

Same happened to me. When asked the Model what Model it is, it will say its sonnet 3.5 no matter what Model I select

1 Like

Yep same here, it’s totally useless now. Hope they will fix this today :crossed_fingers:

1 Like

Hello, I’ve had issues using cursor (on gemini pro 2.5 exp 03 25, with last cursor version and this version) since yesterday too, either it writes the thinking part as normal text and keep repeating it again and again, or it doesn’t parse ``` and writes code as normal text. I’ve also encountered a loop two or three times in a row where it will just continue pasting the same line over and over until I stop it. I’ve also had “connection failed” when trying to continue old chats, but it looks like updating cursor fixed that one.

I spent a good amount of tokens because of these but unfortunately removed the bugged discussions. I specifically created this request for bug reviewing, here we have the “thinking token in normal text repeated over and over” issue: cea99bd6-1d69-401c-b395-b2de69e7979e. (don’t know if you have access to the history, even of the deleted messages because of the revert function, but you will find a good amount of bugged messages in the end of the discussion related to this specific request ID if so)

req ids:
64fb298e-9bd2-4551-81d7-dc95ab7ddb72
f0814240-2b59-4b49-ac28-ffe988ee75e7
81390bd0-76ab-467f-a83f-43319b3fd90d
54a19f5c-8184-4063-8373-11e333471998

And i have many others, but i’ve been branching the chat so i can’t find anymore.

4 Likes