I have a straightforward question about the **Auto** model option in Cursor.
**What exact AI models are included in Auto mode?**
As a paying customer, I often use Auto because it is convenient. But today, Auto is essentially a black box: we do not know which underlying models are being selected, how they are chosen, or how the mix may change over time.
I believe consumers have a reasonable right to know:
1. Which models are currently in the Auto pool?
2. How does Auto decide which model to use for a given request?
3. When the model list changes, will users be informed?
This is not about criticizing Auto itself — automatic routing can be valuable. It is about transparency. When we pay for a product that routes our code and prompts through AI models, we should be able to understand what is running behind the default option.
Could the team publish an official, kept-up-to-date list of models used by Auto (and ideally a short explanation of the routing logic)?
Thanks in advance for any clarity from the Cursor team.
Auto also called Cursor Router or Premium routing picks a model for each request from the models available on your plan. It uses internal benchmarks, quality scores, and current capacity, so the pool is not fixed and can change over time. More details here:
We don’t keep a separate public list of exactly which models are in Auto right now. The set is dynamic and depends on your plan and capacity.
The selection logic is based on task fit and performance, as described in the docs above, but we don’t share the step by step routing algorithm.
If you want full control and predictable behavior, you can skip Auto and manually pick a specific model in the model picker. Then you always know what’s running under the hood.
I’ll pass your request for a public up to date list of Auto models and more detail on the routing logic to the team as feedback. It’s a reasonable ask for transparency. I’ll reply in the thread if there’s an update.
That doesn’t match what I’m seeing in practice. Auto now appears to route virtually 100% of my requests to Composer 2.5, with Opus no longer being selected at all.
This seems to have changed a few weeks ago. Previously, I could see Auto switch to Opus for more complex tasks, while using faster models for simpler work. Now, even difficult tasks remain on Composer 2.5.
If the model pool is still dynamic, could the team clarify whether Opus has been removed from Auto or whether the routing thresholds have changed?
This seems to correlate with the new types of auto that were only introduced with the Teams and Enterprise plans.
Thanks for sharing what you noticed. I’ll answer this separately from the general transparency topic.
The model pool in Auto is really dynamic. What you get depends on your plan and current capacity, and it can change over time. So the fact that you used to see Opus on harder tasks, but now you see Composer 2.5 more often, doesn’t automatically mean Opus was removed from Auto. We don’t publish a per-factor breakdown of the routing thresholds, so I can’t confirm or deny a specific change just by guessing.
To see what’s actually happening in your requests, please send:
A couple of Request IDs where you expected Opus but got Composer 2.5 (Chat menu top-right > Copy Request ID). Note: if Privacy Mode is on, the Request ID might be unavailable.
Your plan (Pro / Team / Enterprise).
Your Cursor version and which mode you’re using (Auto, Auto Cost, etc.).
With that, I can pass it to the team to check the specific routing.
If you want predictability right now, you can avoid Auto and manually pick the model in the model picker, then you’ll know exactly what’s running under the hood.
Hey, thanks for the request. It makes sense. If an org limits the list of allowed models, you’d want Auto to pick only from that allowed set, not switch to a model outside the allowlist.
Right now, Auto picks from the pool available on your plan, and a user-selectable allowlist for Auto isn’t supported yet. I’ve shared this as feedback with the team.
There are a couple related requests on the same topic. Feel free to check them out and add your vote or details, it helps the team see demand:
Workaround for now: instead of Auto, pick a specific allowed model in the model picker. That way you’ll definitely stay within your compliance list. If there’s an update on this, I’ll reply in the thread.
Thanks Dean, here’s a complex task that was rooted in what I believe was Composer 2.5. I can tell, as it was extremely fast.
7efdd6eb-e7f9-4919-97a9-5515a74a793c
Also, I am on the legacy plan pricing
Can you confirm or deny whether Opus is included in the pool of available models for auto? Now that you have your own Grok 4.5, which is close to Opus’s capability, wouldn’t it be much more cost-effective for you guys to just use that instead of Opus?
Thanks, I got the Request ID. The fact that it finished very fast doesn’t prove which model was used, so we’ll check the logs for this specific request.
A couple quick questions so the team can review the routing more precisely:
Your Cursor version and which mode you used, Auto or Auto Cost.
If you can, please share 1 or 2 more Request IDs from similar complex tasks where you expected Opus. With a few requests, it’s easier to see the routing pattern than from just one.
One important note: on legacy pricing, the model pool in Auto can be different from the new one. It depends on your plan and current capacity and it changes over time, so it’s not always correct to compare directly to how it worked before. But with the Request IDs we can see what was actually selected for your requests, and I’ll pass it to the team to review.
If you need predictable behavior right now, pick a specific model in the model picker instead of Auto. Then you’ll know exactly what’s running under the hood.