Hey, thanks for the report. From the screenshots I can see usage for Sonnet 5 and Gemini 3.1 Pro, even though you only selected Grok 4.5 and Composer 2.5.
What’s happening: when the main agent delegates part of the work to subagents via the Task tool, the subagent can pick a different model than the one you chose for the session, sometimes a more expensive one. This isn’t anything in your setup, and you didn’t do anything wrong. It’s a known issue we’re tracking. I can’t share an exact ETA yet, but I’ll update the thread when there’s news.
A few things that might help right now:
- In the CLI
cursor-agent, there isn’t a full model catalog for subagents yet, so you can’t directly pin the subagent model there. - If you also use the desktop app, you can set the model for subagents/Task in settings, which gives you more control.
- To gate subagent spawning, you can look at the
subagentStarthook Hooks | Cursor Docs. It lets you allow or block subagent creation, including by model.
Related threads if you want to follow along:
Separately about quota usage: if you want us to check your specific account, what was charged and what we can do about it, email [email protected]. Our accounts team handles that. Let me know if you still have questions about the subagent behavior.