So there’s a new “Auto” mode called “Premium”.
Will using the Premium setting stretch the limits better than simply setting and forgetting model to “Opus 4.6”?
5 Likes
Yes I does, you going to use significantly less credit. As I experience like 1/3 compared to using Opus 4.6 only with minimal Sonnet (that is what I used to do). Although It worth noticing that the “Premium“ mode 99% of the time uses GPT 5.3 Codex, which is a much cheaper model than Opus. So basically what they did here, is they wrapped the Codex model behind this premium mode basically, as the mode itself does not feel good enough to truly determine which model is the best for your case, so you will really likely will end up using Codex and and for which you will pay the basic API pricing. I am not saying it is not good, but definitely does much sloppy research and planning the Opus usually does and the code quality is really not the same for sure. On the another hand it is cheaper. I am not 100% sure if you would not pay the exact same amount just with using 5.3 Codex by itself.
1 Like
Great feedback! Thanks to you, I’ll keep using Opus for the AGI level stuff, and consider Premium for simply “difficult tasks".
Of course, unless Premium is actually at least 30% cheaper, I prefer knowing which model is working under the hood. I know they are well grounded and all, but they still have their personalities and unique vibes.
So the question still stands, if anyone has benchmarked the OpenAI model against Premium mode?
We need voluntary Cursor benchmarkers!
In the usage tab, after each request if you use the “Premium“ model it publicly disclosing at which point which model is getting utilized, and you are getting billed for that model. So as I said before, the “Premium“ is just a wrapper that “intelligently“ decides which API model they should use for the given task (for me most of the times 5.3 Codex ).
Hey, good question.
Auto and Premium are two different modes with different billing pools:
- Auto uses a separate pool with a fixed price per token, which is much cheaper. It’s good for everyday tasks.
- Premium routes requests to specific API models and charges from the API pool at that model’s rates. So if Premium picks GPT-5.3 Codex, you pay the Codex API rate. If it picks Opus, you pay the Opus API rate.
In practice, as many users have noticed, Premium usually routes to GPT-5.3 Codex, which is much cheaper than Opus 4.6. So yes, Premium is usually more cost-effective than set and forget Opus 4.6, but the quality will differ because Opus is stronger for harder tasks.
If you’re on Ultra and want maximum quality for AGI-level tasks, pick Opus 4.6 manually. If you want to stretch your budget, use Premium or Auto.
More details on pools and pricing: Models & Pricing | Cursor Docs
In the usage dashboard, you can see which model Premium picked for each request, so you can track how much you’re saving.