Composer 2 just redirects to GPT-5.2?

Support Help
,
1

image
image952×560 13.9 KB

Is Composer 2 at Identity Loss or is it just GPT-5.2 but renamed in the UI. I know it is powered by Kimi K2.5, but that’s not what I mean. I mean I use GPT-5.2 Xhigh frequently, so did the model selector just stop actually changing :face_with_spiral_eyes:

image
image982×544 12.5 KB

3

Hey, from the screenshots I can see that Composer 2 Fast is selected in the picker, but the model introduces itself as GPT-5.2. I get why that looks suspicious.

This isn’t a routing bug. Models aren’t reliable at identifying themselves. If you ask an LLM “who are you?”, it might call itself anything. That’s a hallucination, not a signal of which model is actually handling the request. Composer 2 doesn’t have to “know” it’s Composer 2.

The picker in the bottom bar is what determines the real routing. If it’s set to Composer 2, the request goes to Composer 2.

If you want to confirm the models are different, try giving the same coding prompt to both and compare the style and response speed. The difference should be noticeable.

4

I find this interesting. In fact, it intrigues me how this only happens with Composer. Furthermore, it takes the identity of any model I often use. When I heavily use Opus, it takes that identity, when I heavily use GPT-5.2, it takes that identity (only Composer). All other models are fine. Lastly, I don’t feel auto actually picks. it uses the same model for like 50 tasks in a row, even though some are very complex and some are very minor. I once had it use Opus 4.6 for a week, and then Codex 5.3 for a month. I don’t know if cursor is defaulting/hardcoding these. Can you please clarify on all three of these points? Thank you.

A Loyal Cursor User

5

Hey, here’s a point by point breakdown:

  1. Why Composer uses other models’ identity. This isn’t because you use them a lot. Cursor doesn’t track your model usage and mix it into replies. It’s more likely that Composer 2 was trained on data where popular models like GPT and Claude are often mentioned as assistants. When you ask who are you, the model generates a statistically likely answer, not a fact. Models like GPT-5.2 or Opus tend to have a clearer system prompt for identity. Composer has less of that binding, so it can drift. That’s hallucination, not routing.

  2. Auto really does switch models. It routes based on the type of task and available capacity. It’s not rotating just to rotate. If one model consistently handles your requests well, Auto can stick to it for a long time, and that’s normal. If your requests are similar even if they feel different in difficulty, the router may decide one model is enough.

  3. There’s no hardcoding. A specific model isn’t pinned to a user or account. Auto logic is dynamic.

If you want to confirm Composer 2 and GPT-5.2 are actually different, give both the same prompt and compare style, speed, and tool call approach. The difference should be noticeable.