I’ve done a lot of tests and experiments, and I’m almost 99% sure that it is indeed a model — a specific model trained for Cursor, very fast, and able to use Cursor’s tools extremely well.
Unfortunately, it’s clear that GPT-5 isn’t great for UI, and AUTO — the “actual model” — is better.
If it’s clear to you otherwise, then sorry.
danperks, a Community Developer said the following
Auto is designed to use a premium (1-request cost) model while ensuring a stable, reliable output! It’s best for people who don’t care for learning the intricacies of each model and also for working around model downtime and outages with the inference providers.
Due to how it works, it does not respect the list of models you have enabled, but should give you a consistent level of intelligence each time.
So it couldn’t possibly be a model itself but a selector based on availability.
why does auto tell me he is claude 3.5 sonnet
It is related to his system message and the preparatory materials.
And it’s not always what he writes to me; he writes other things.
I know that’s what they say, but I can tell you it’s not true.
I wrote a long post about it.
In short, if it were true, they wouldn’t mind showing you every time which model it is.
Of course, it’s always possible that I’m wrong, but I use this model countless times, and it’s the same model every time.
@Kody_Fisher when will this take affect for the cusor-agent CLI ?
* she
Bring back reapply….
GPT-5 has been counted toward the limit and is within the free period
should be at least $1000 since its an x10
How you got requests based pricing?
I wrote an explanatory post about this. I know that’s what it says, but the question is whether what’s written is what actually happens in reality.
I believe it isn’t!
Check, try it yourself, and you’ll probably end up thinking like me in the end.
I’m 100% absolutely certain Auto does what they say it does. I’ve seen model switches. Don’t you notice how different models produce different kinds of output? Auto absolutely switches between models. You can see the difference in behavior. One model thinks a certain way, maybe taking a long time in big chunks of thinking, while another thinks in little short stints before doing work, then thinking a little bit more. Some models don’t think at all.
How the heck haven’t you noticed model switches when using Auto mode??
I have no way to prove it.
I haven’t seen what you’re saying.
What I have seen is a very good model for UI better than GPT-5!
Amazing at finding things quickly.
The only model that really works well with the terminal, at incredible speed, in areas where every other model gets stuck and takes forever.
If it’s clear to you otherwise, then I don’t know what to say.
Of course you have no way to prove it. I suspect you are not a very high volume user, as you don’t seem to have pushed Cursor to its limits. Auto is a mode, not a model. Its design is to give Cursor the ability to switch models on you. This is why they were allowing users to use it for free…it BENEFITS CURSOR, not the user!
By allowing Cursor to switch a model on you, which could, can and very much will, reduce the quality of your results if you get a slow or poor quality model, Cursor can save money. That was the cost-benefit they were trying to capitalize on. Thing is, compute aint free. Cursor’s auto mode is a way of distributing load, to try and share the cost around the various providers, rather than concentrate those costs into one basket (i.e. Anthropic, which sadly, seems to be having trouble meeting their compute needs, as it is the single most used model for AI generated code with Cursor, CoPilot, Cascade, etc.)
Simple fact of reality is, AI computation is not free. In fact, it has become the most energy-hungry technology we’ve yet developed, outside of perhaps blockchain. Everyone who is demanding that they get unlimited tokens for $20 a month is literally delusional. No company can actually deliver that, and survive.
Do people want Cursor to survive? Do they want Anthropic to survive? Windsurf Cascade has had problems too. Their top execs were poached, and the product and dev team sold, because they weren’t surviving. Cascade is now behind Cursor. Who knows if it will be able to catch up. The landscape here is, compute aint cheap. We are literally building entire power plants, nuclear power plants, to supply the power needs for LLMs. Those costs have to be recouped, the energy costs have to be recouped…
THAT is what Auto MODE is about in Cursor. Its not a model.
How can I measure it, so I can tell you for sure?
Right now, my total usage for this month is at…
If that is your usage, then you most certainly have experienced auto mode model switches, if auto is what you use.
In any case, Auto is a mode, not a model, that is evident from the docs, and evident from the usage experience: The model with auto CAN and WILL change on you, when Cursor’s systems determine that the model you started with is overloaded (what their criteria is, I cannot say.)
I’m not going to debate the topic further, as its a ridiculous debate. 
In any case, I enjoyed talking with you.
I read everything you wrote so far, and you really do sound experienced and like an expert.
