My main agent was using Cursor Grok 4.5 and invoked multiple subagents during a task. I had explicitly configured all subagents to use Composer 2.5, and one of them did use Composer 2.5 as expected, but another unexpectedly ran on Sonnet 5 High without my approval or any warning. This unintended model selection caused my usage quota to decrease significantly. Please investigate the model routing and usage associated with Request ID c95984a2-d2a8-4ce0-9efa-31ef71103f7a, clarify why the configured subagent model was overridden, and consider restoring the quota consumed by the unintended Sonnet 5 High execution.
Steps to Reproduce
Configure the main agent to use Cursor Grok 4.5.
Explicitly configure subagents to use Composer 2.5.
Run a task that causes the main agent to invoke multiple subagents.
Review the models used by the invoked subagents.
Expected Behavior
All subagents invoked by the main agent should use Composer 2.5, since that is the model explicitly configured for subagents.
Hi, there! This is expected behavior, the composer 2..5 model you configured is just for “explore subagent”, not for all. You can prompt it if you want use composer 2.5 for subtasks.
This happened to me too (though i didn’t explicitly configure the subagents), and it’s very upsetting that a prompt intending to use the first-party models used up all of my api credits in one prompt.
Hey, thanks for the detailed report and screenshots, they make everything clear.
The setting you changed (Explore Subagent Model → Composer 2.5) only applies to the Explore subagent, not all subagents. You can see that from the label in the settings, model used by the Explore subagent for initial research. Your second screenshot also shows the Explore subagent did run on Composer 2.5.
The second subagent, Implement remaining Android features, is a regular Task subagent, not an Explore one. For those, the model is picked by the parent agent, and that’s why it ended up using Sonnet 5 High. There isn’t a separate UI setting right now to hard lock the model for non-Explore subagents, so Congzhi is basically right.
What you can do right now to avoid this happening again:
Turn off Sonnet 5 and any other premium models you don’t want subagents to use in the Models list, then subagents won’t be able to select them
Keep the parent agent on a cheaper model
Watch the subagent cards and cancel early if a subagent starts with the wrong model
Not having a way to restrict the model for all subagents is a known item and it’s on the team’s radar. I can’t share an ETA yet.
If I get an update on subagent behavior, I’ll post it here.
Thanks for the clarification, that makes sense now. I didn’t realize the Composer 2.5 setting only applied to the Explore subagent, while regular Task subagents could have their model selected by the parent agent.
I’ll disable the premium models I don’t want subagents to use for now. That said, I still think this behavior is quite easy to misunderstand from the UI, especially since it can lead to unexpectedly high quota usage. It would be really helpful to have a global subagent model restriction, or at least a clear warning before a Task subagent starts using a premium model.
If possible, I’d also appreciate it if the usage from Request ID c95984a2-d2a8-4ce0-9efa-31ef71103f7a could be reviewed, since the quota impact was significant. Thanks again for the detailed explanation.
Glad it’s clearer now. I totally agree with your UI feedback, that you can’t lock the model for regular Task subagents, and there’s no warning before running a premium model. The team already has this on their radar, but I can’t share an ETA yet. If there’s an update on subagent behavior, I’ll post it here.
About the quota for Request ID c95984a2-d2a8-4ce0-9efa-31ef71103f7a, those requests are handled by the team at [email protected]. Email them with this Request ID and they’ll look into your specific case.
Thanks for confirming and for pointing me to the correct support channel. I’ll email the team with the Request ID. I appreciate the clarification and hope a global model restriction or a warning before premium models are used by Task subagents can be added in the future.
