Cursor 4.7 "Auto" model selection

[non-npc]

In 4.7 they introduced a “Auto” model selection when you start a new project. This is terrible because it switches model on you if there is network congestion or high model traffic/usage.

So Mid-project or any point really, it switches model on you which causes it to lose context of the project. You can be coding with Claude 3.7 and suddenly it switches you to GPT O-1 mini and it has no idea about the project or anything that you were working on and starts breaking the entire project.

So go to the Cursor settings and make sure to unselect all the models you do not want it to use. This way when you start a new project it will only “AUTO” from the selected models.

[binatheis]

Exactly, I didn’t like this addition and the worst of all is that you deactivate it, close Cursor and when you open it again the Auto mode is activated again. If I decide to use a specific model is because I need it and that’s it, nothing should guide my choice, the change is only made at my discretion.

[gprethesh]

i don’t understand why would they behave like a corporate they would try to hide everything. they have hidden now @codebase we cannot mention this. i don’t understand how can a company treat its users by hiding so many things. i think someone should file lawsuit for hiding many things.

[parisneto]

I agree this was a bad move the way Auto Model was implemented.

Removing the transparency ■■■■■. I am ok with optimizations, but this is probably true to Cursor owners saving money on LLM costs. For me is degraded performance, experience.

It’s happening frequently ( at least once per 3h code session ) when I start to see hallucinations and low quality or incomplete responses, I am loosing time, the I note : it reverted to AUTO even after I choosed a model many times and disabled. Likely it happens in a new chat windows or after restarting cursor, in a non transparent or expected way.

It really seen like a feature designed by a CFO that doesn’t code and only look at numbers on a spreadsheet and will get a bonus to invent how to save money , and of course CFO performance is measured by money only, sorry for the joke, but I bet you all seen such corporate game, there are good ones out there and we should praise and care.

But hidding models to avoid risk, lack of transparency what is running , silent switches of models is REALLY BAD, destroying trust, experience and satisfaction of a paying customer.

and I did not opted for this, I was nice to test as a Fan of the old Cursor days and I pray ( and write here ) for them to listen and pivot.

there is a lot of smart people suggesting heuristics to use models for specific tasks, that would require transparency, choice and some sort of return to the user, not only savings for the company.

MOST IMPORTANT : ANYONE FOUND TO DISABLE THIS COMPLETELY ???

[danperks]

When you have ‘Auto-select’ enabled, Cursor will pick a Premium model for you to use. This is usually Claude 3.5 or 3.7 - Cursor won’t put you onto a thinking model like o1-mini!

We have previously pushed people to Auto when we’ve struggled for capacity from Anthropic, to help keep things running smoothly, but I’m hopeful this will happen decreasingly often as we improve things further around this.

[non-npc]

on the surface this seems like a non issue but it causes the new agent to lose all context of the project being worked on

[parisneto]

Dear Sir @danperks (and CC: cursor team), I do hope all is well first,

I am consistently learning from experience this week, that when I see a lot of lower quality code, solutions that ignore my prompts and change comments (not code) and create bugs, that I was unaware of being in auto mode.

And that is very frustrating :
1 - there is no transparency of which model I am using, so I never learn from experience which are good and bad for my use cases
2 - I disabled repeatedly auto mode, and was unware when then change to auto happened, this could affect a 3-4h code sprint requiring reverting a git commit and starting again.

I would much rather know there is a capacity problem ( I am aware of them, I know it’s not Cursor fault ) than being diverted to a unknown model, that only proved to be worse , spending, I mean wasting many hours, fast requests listed as Default (ZERO TRANSPARENCY on billing also).

And I am a fan and supporter of cursor, you, @deanrie and this community, but I really feel this non transparent, auto switch is causing a lot of harm, reducing my value and productivity using cursor. I already said in anothe post, seems like a “bad CFO” invented feature to save money, that never coded and will get a fat bonus, not something carefully crafted like most we see here. And I see the posts for ideas using Auto with heuristics and agree it’s the future ideal world.

But let me be clear, I feel disrespected and robbed by this non transparent auto switch and lower quality experience. This is really bad and not in line with all the great work Cursor team has been pushing forward.

Please consider a FAST FIX to disable entirely this maddening and worst update ever for Cursor. If there is anything I can do already, it’s nowehere to be found ( with my skills ). I would greatly appreciate to learn how to get rid of this entirely.

Image references :

• billing
  

  

• great fix for AI , change the comment to say the same thing and call it a fix :
  

  

• Look at the lousy problem created in Auto mode , confession of a bad decision (context: moved some settings to a central config file, and instead of fixing the methods, names to use it, just add to the code and criticize the solution hardcoding with errors, that I learned and had to fix later :
  

  

one comment, for many months I review every change line by line, because the quality for all models varies a lot. Sometimes in a session that you are making simple things in fewer files/context I take some risk with models I trust like Sonnet 3.5 or 3.7 but then the non transparent auto switch caused this , charged as much as premium fast requests, so I rather see a cannot reach anthropic than this, every day , every time .

and you mentioned :

We have previously pushed people to Auto when we’ve struggled for capacity

this is happening this entire week for me.

I hope you read it all. Have a nice friday and weekend !

Cursor information :

( There are currently no updates available.)
Version: 0.47.8
VSCode Version: 1.96.2
Commit: 82ef0f61c01d079d1b7e5ab04d88499d5af500e0
Date: 2025-03-18T06:55:51.040Z
Electron: 32.2.6
Chromium: 128.0.6613.186
Node.js: 20.18.1
V8: 12.8.374.38-electron.0
OS: Darwin arm64 24.3.0
+
Macbook Pro (m3pro) on Ventura 15.3.2 (24D81)

[parisneto]

Additional confirmation :

even being like this, I was silently pushed to Auto frequently and cannot identify when ( new chat, longer session, reopen cursor ) but it’s not respecting my setting of Auto-Select turned off.

[T1000]

Only on new start it would happen. Their team was informed.

You can also set for each mode a specific model for now.

[parisneto]

@T1000

new start you mean opening the Cursor IDE or workspace ?
for me it happens even without reopening Cursor, in the middle of a 3hr coding session.

Where I can do this that you mentioned :

You can also set for each mode a specific model for now.

when you say mode you mean ask/agent ?

I double checked the UI and Cursor Settings and was unable to find.

Can you clarify and compare your cursor version ?

Version: 0.47.8
VSCode Version: 1.96.2
Commit: 82ef0f61c01d079d1b7e5ab04d88499d5af500e0
Date: 2025-03-18T06:55:51.040Z
Electron: 32.2.6
Chromium: 128.0.6613.186
Node.js: 20.18.1
V8: 12.8.374.38-electron.0
OS: Darwin arm64 24.3.0

on Mac Os Sequoia

[T1000]

Ah i see. you have 0.47 and im on 0.48 beta version which will roll out soon i think as stable.

Yes on 0.48 you can set more things.

For me it happens only on restart of Cursor.

[non-npc]

It event switches to AUTO mode even if you have specifically set the model to the one you want. I would rather be put into a timed queue than deal with the AI losing context of the project and having to waste time teaching it the project again. If Cursor really wants us to have this then perhaps they need to add a way to make the AUTO model chosen have context of the project so the user does not become frustrated with why the AI model suddenly acts like it has no idea what is going on even after hours or days of work

[sailxjx]

Yes, it’s really annoying. I always have to manually switch back to Claude. Every time cursor automatically switches to the auto model, the modifications I get are very poor (which is also why I noticed the switch to auto), and I’m not clear on what model they actually used.

[ppt]

Removing all other models from the model list, leaving only the models we need (like Claude 3.7), is not working.
I tried doing this, but actually when I asked it “What model are you?” it answered me Claude 3.5 Sonnet.

[DJFormica]

@non-npc

I would rather be put into a timed queue than deal with the AI losing context of the project and having to waste time teaching it the project again.

This. I don’t mind waiting for an answer if I know the answer will be consistent with the work I’ve been doing and the answers I’ve been getting. I really like Cursor, and I just signed up for a paid subscription as I find it superior to its competitors. But this “auto” change dramatically impacts Cursor’s usefulness.

I get the need to reduce costs, and I’m even sympathetic to it. But don’t abstract the activity from the user. Give us the option to wait, or, ask us if we want to wait or switch to a different model. Please do not make it a black box activity like it currently is. Losing the context and having to start over is antithetical to the entire point of AI coding assistance - saving time.

[duyvinh229148]

fking hate this Auto-select ■■■■■■■■■■■■■, ■■■■ so annoying, I’m so angry that I have to find this post and comment. All I want is claude3.7, that it. Ruin my vibe coding

[robshox]

Is there a way to see what model has been selected? I get the idea of auto mode but using AI is like working with different co workers. If you switch the model on me it’s like switching my pair programmer from simon to francis. I understand the intricacies of a model and prompt in a particular way for each

[akali95]

I can confirm this still happens in 0.48.x.
This is actually really annoying, and frankly quite insulting to paying customers. The model randomly changes throughout the day from my selected model to the auto model selection and I have to figure that out after I’ve wasted a premium request on a model that I do not wish to use.
I don’t WANT to use a random model, I want to use my model of choice, if there are capacity issues or anything else, then I am fine with having to wait. But this absence of transparency and switching models around randomly (and CHARGING a fast request for a model the user was not aware he was using) is not acceptable and borderline fraudulent behavior.
I like Cursor, I really do, but this is not the proper way to treat your customers.