It keeps flipping back to Auto, and it’s making even more mistakes. Stop making the agent flip to Auto—this is a real pain. If I don’t notice you forcing the flip, suddenly everything gets even worse. This is really bad practice to force this on users. If I wanted Auto,
I would choose Auto. I do not want you forcing Auto on me just because you think I’m too dumb to choose it. I want it to remember which model I’m using.
just arrived here to mention the exact same thing.
Agree with this, not sure why it always switch to Auto when I launch cursor.
Is there a good explanation somewhere as to what Auto does?
I wouldn’t mind cursor auto selecting cheaper models when it can… but want to know how that would work?
Hey, so Auto always picks a premium model to use, but has some intelligent routing under the hood that should make it much less at risk of connection errors due to model outages or capacity issues.
However, a lot of uses still likely to manually choose their model. Usually, we’d be more than happy with this - user choice is never a bad thing - but on occasions where we are running out of capacity for the Claude models from Anthropic, we sometimes fallback users to the Auto setting, so they can be temporarily routed to other models and we can keep everyone working smoothly.
For most people, this usually isn’t an issue, but for more advanced Cursor users who know the intricacies of certain models, this can be somewhat disruptive. We by no means want to have to do this, and are working hard to mitigate these capacity issues with Anthropic and our other LLM providers, but at least this way peoples workflows don’t come to a halt if a model is suddenly unavailable.
We hope not to have to use this option as often as in the past, as our capacity stabilizes moving forward.
agree
Agree. I have never set the model on AUTO, but I keep finding it set there.
No Dan, this is not a solution—this is a bad idea, full stop.
This has repeatedly ruined my codebase, regardless of whether it’s a premium model or not. The flow gets wrecked when you decide to flip it to Auto. If it can’t be done from Anthropic, then STOP—don’t ruin my codebase, forcing me to restore from Git and go around in circles, only to realise it’s in Auto again.
This downgrade shouldn’t be forced on so-called “advanced” users. It doesn’t help—it’s cost me hours and hours. You say it doesn’t disrupt workflows, but it’s wrecked mine for seven days with zero progress. It’s a mess. It can’t even solve basic problems now, can’t remember anything, and to top it off, it switches to Auto and starts coding exactly what it was previously told NOT to do, over and over again.
Whatever you’ve changed isn’t right. Just over a week ago, it was flying—coding complex setups perfectly. Now it can’t even pass a variable between pages without breaking, constantly checking the database after every command and wanting to restart Docker every time. I could go on, but it’s clearly in bad shape—and this Auto nonsense only makes it worse.
I’ve had to turn off every other model just to stop it from falling back. But it feels deeper than that—like 3.5 is being masked as 3.7.
It can’t solve the same issues that 3.5 struggled with in 3.7 settings.
It codes like 3.5.
It feels like 3.5.
Maybe it is 3.5.
Please get this sorted out or i will have no choice but to leave this it 's too far behind and anthropic needs to be negotiated with to give cursor properly allocation levels, not this constant breaking and unreliability. I would not be surprised if there are peak and off peak times to tell cursor users when they can get the most out of the software too, like maybe overnight usa sessions is quite i don’t know just a thought.
We’re happy sticking with what we know. Why use a model that generates buggy code and waste time and money fixing that mess? Which guy on the Cursor team came up with this feature? Did you guys let a sales guy make decisions instead of an engineer?
Interesting motivation for keeping ‘everyone working’, bad practical implementation.
Maybe adjust that to keeping ‘everyone productive’. Fixing bad AI slop is wasteful.
Even noobs can tell the difference between the quality of model output.
2 weeks ago I created a post to memorize that experience 
100% agree. I ended up just removing 3.5 from models in settings
Threat-Driven Development: Thanks for the trauma, Auto Mode
Claude 3.7 made me believe in AI again. It was zen. It was flow. It was love.
Then Cursor dropped “Auto model selection” and boom — I was back in the dungeon with Claude 3.5. No warning. No badge. Just broken dreams and a deranged assistant who forgot the rules of the conversation it started 3 minutes ago.
I didn’t realize what happened at first. I just felt the return of rage. Full-caps Russian swearing. Threats of unprintable violence. Productivity imploded.
Eventually I figured it out: Auto = Surprise downgrade. So I manually switched to Claude 3.7, and like magic — the dev heavens opened again.
So now I do a little ritual every morning:
Open Cursor.
Switch model to 3.7.
Whisper “Goose-fraba” while praying they don’t sneak me back into ■■■■.
Request:
Give us a ■■■■ model label.
Don’t bait-switch. Just ask.
Stop triggering my coding PTSD.
Still love you, Cursor — but ■■■■, that was rude.
– A traumatised dev
The issue is that it used to behave differently. Just a couple of weeks ago, Auto was consistently defaulting to Claude-3.5. I’m sure it’s picking a premium model now, but the inconsistency already caused enough damage — the “Auto-model” feature left some real scars.
That’s exactly the kind of Cursor “feature” that makes me clench every time I see a “restart to update” notification.
Very agree, do not default to auto, the customer decides what large model
At a bare minimum, it should tell you which model it picked while it is working and label the responses accordingly. I tactically switch models all the time based on what kind of response I want, but a key part of that strategy is knowing what I’m working with. Auto-picking isn’t a deal breaker for me, so long as I know how to interpret what it’s doing.
For example, I know 3.7 is smart and capable but also easily distracted by shiny objects, and I need to know to interrupt it when it starts chasing some random thing that caught its attention. If Cursor is flapping between models, I can’t be sure which behavior to expect until it’s too late.
Very bad idea.
I’d rather see “unavailable”, and choose an alternative and adapt my prompt accordingly.
This is why I am using a previous version of Cursor (0.45.17)
It has its own bugs, but still better than the “new features”
