Hey @Mihai_Cracan
The difference isn’t native vs custom model as such. Cursor decides which prompt and tool set to use by looking at the model name. If a custom model’s name contains one of our built-in model ids (for example anything containing gpt-5.6-sol), it gets the same GPT prompt and the same apply_patch based editing tools as the built-in entry, even when it’s served through your own API key or base URL. If the name doesn’t match a built-in id, it falls back to the generic harness with the StrReplace/Write tools you saw.
That’s why the naming matters more than it looks. A name like custom-5.6-sol won’t match because the check is for the literal gpt-5.6-sol substring, so it gets the generic setup. Renaming it to something like custom-gpt-5.6-sol-medium would put it on the GPT harness. Incidentally, the my-custom-gpt-5.6-sol you mentioned in your post would have matched.
We will share more in the coming weeks around continuity of access to OpenAI models. That news probably won’t be coming from me in a forum thread.
Stay tuned.