Auto is now for efficiency, while Premium is for quality.
Is there any list of which models can be picked up by auto and which by premium? I believe both of these options are for less technically knowledgeable people, who do not care about the model, just the results. But it would be great if such information was in docs.
Also thank you for adding separate pages for some of the models into documentation, it is great!
It seems I’m still on the old UI because I don’t see the Auto and Premium options. For now, if you want to manage models, I’d use Kimi K2.5 as a daily driver since its price per token is good and it’s also quite fast. I use it as my daily model, and when something is hard, I switch to Codex 5.3 (I haven’t tested 5.4 GPT yet)
Draws from api pool, so it automatically chooses different API models (non composer or auto?)
@valentinoPereira Premium may be like Auto but it routes you to models that charge API Usage. Their needs to be more documentation. So in a way this could make it simpler if you don’t want to choose models ever.
Update!
Not sure if this is new or was there since the beginning but you can see which model is being used by this new “model selection” in your dashboard:
Sadly, auto still does not show which model it uses.
It is because you are actually being charged for that specific model chosen by Premium, so you need to know what it is (ie Opus vs gpt 5.3 codex would be very different charges). Auto charges you a flat rate regardless of what model it chooses (so its hidden), but Premium is just selecting the model for you and you are still paying that model’s api rate. It basically is selecting the model for you as if you did yourself from dropdown model list. Auto and Premium do very different things, and I think Cursor is just going to confuse people by putting Auto and Premium next to each other.
Old.
Can program.
UK CS educator.
First computer 1983
Has a Masters Degree in AI (If I pass it in 2 weeks time)
I spotted this premium routing option about half an hour ago, which brought me here as I want the Jerry Springer Final Thoughts section of my dissertation to be up to date and forward facing. My dissertation is based on the concept of AI curation.
I love auto. I was burning through ultra tokens in fewer than 2 weeks for my work and research so it has saved me a fortune and is very competent. I don’t mind that it is not transparent. The output is better than all of the cheaper models I have used so I suspect I am getting curation + discount on more expensive models for a solid price and consistent performance.
Give the system prompt team a pay rise.
This new feature seems to be giving me an AI curation of the prompt and then the API call to a frontier model which is transparent in the usage section of my dashboard (confirmed) but is not charging me anything on top. That is an epic deal.
The only thing is, and if you have read this far I trust you agree that I have sufficient experience in this area, the response it gave back to me is not in the same style of the model I have been charged for.
Looking carefully at the routed output it gave me, I think I can see what is actually happening. The tool call narration, which is short, precise, and has no personality, reads like Composer doing the agentic traversal. The synthesis at the end switches register entirely into something that is unmistakably GPT IMHO (in this instance). If that is what Premium routing is doing, it is genuinely impressive. Composer curates the context and then GPT, or whatever it selects, closes it out, and you get output that neither would produce alone at that quality.
The premium router is always picking “premium (gpt-5.3-codex)” for my current tasks, according the dashboard. This is not unreasonable given what I am doing.
Is anybody seeing different frontier models being consistently selected?
So, for those who bought Cursor before September, Auto was completely free and included Premium models. Now Auto only uses lower-tier models, and Premium models are no longer part of it. So the only way to maintain the same level of quality is to pay more or switch to a different pricing plan. Bonkers.
I actually know that. I use it myself. But the lower-end models should at least have unlimited usage. This is a bit unfair. Not everyone uses the higher-end models.
I have not read that Auto is being “downgraded” but one could assume that with the introduction of the Premium feature. However, I think there is some confusion. Premium is not like Auto at all. While Auto charges a flat rate for requests and behind the scene routes them to different models, Premium is just an auto model selector. The requests using Premium are the same as if you had chosen that premium model yourself. It’s for users who don’t want to determine which model to use, but rather have Cursor interpret and decide. So Auto may not be being downgraded, because using Premium is technically not any different than what we already had with manually selecting our models. Auto as a product should not be impacted, in theory. Also I doubt Auto was routing to premium models anyways. It was probably routing to Composer and GPT-5 most of the time; their API prices are similar.
There is no unlimited anymore since last September, except those grandfathered into it (at most for a year). The exception is GPT-mini which last I checked does not count towards usage limits.
Why should lower end models be unlimited, how is it unfair for Cursor to not give us unlimited free stuff?
It’s doing the same thing for me. I’ve been doing a lot of work with it, and I haven’t been able to get it to switch to any other model other than the GPT 5.3 codex when using the premium router and Cursor.
