I was using Grok 4.5 - High (not fast) and it decides to spawn a subagent Composer 2.5 (Fast) which I configured Grok and Composer for non-Fast option. It consumes lot more token which I do not want.
Steps to Reproduce
Technically it depends to the AI model, so i do not know how to reproduce
Expected Behavior
For e.g. if I configure the Composer 2.5 or Grok 4.5 with non-Fast option, i expect the system must respect to choice. It shouldn’t be allowed to select expensive options on its own.
To explain what’s happening here: subagents don’t automatically inherit the model from your main chat. By design, the agent picks whichever model it judges is the best fit for each subtask it spins up, and that can include Fast variants even when your main session is on a non-Fast model. So this is expected behavior rather than your model selection being ignored, though I understand it’s surprising when it shows up in your usage.
A couple of ways to get more control over this:
Configure the Explore subagent. In Cursor Settings → Agents → Subagents you can set the model for the Explore subagent explicitly. Note: this only applies to Explore. Other subagent types the agent creates aren’t covered by that picker.
Disable models you don’t want used. The stronger option is to disable the unwanted models in Cursor Settings → Models. The agent chooses subagent models from the set of models you have enabled, so removing the expensive ones from that set prevents the agent from picking them.
You can also add a rule (e.g. “subagents should use the same model as the main session”) as a softer nudge, though rules guide the agent rather than enforce it.
Hope that helps, and let me know if you have any follow-up questions!
Same issue here. GPT-5.6 Sol is using Sonnet 4.5 non-thinking for subagents even though I did not select or want that model. Sonnet 4.5 non-thinking is old and should not be an implicit fallback.
Cursor should either disable it for subagents by default or give us an enforced per-subagent model allowlist. Having a model enabled in the picker should not mean an agent can silently spend usage on it. Respect the selected model, or ask before switching.
I understand the purpose here, but I still don’t understand why the design doesn’t forcibly restrict the subagent selection for the model when there are models that specifically selected/configured that way. As you said, if I disable models in the settings, the subagent can’t select those models, which makes sense. However, on the other hand, even if I select Composer 2.5 Non-Fast as the main model, the model can still select Composer 2.5 Fast. What’s the logic here?
Yes, they could be the same model (just faster), but as you see, I didn’t select it, and the subagent shouldn’t either. To be short, the subagent picker must respect the Fast/Non-Fast model choice.
While fixing this with a rule is actually not that bad. I still would also prefer if you guys could add some configuration so I can unselect active models for custom subagents spawned by agents.
Starting a task with Cursor Grok 4.5 to find out afterward that 80% of it was delegated to GPT 5.6 Sol is just a really bad user experience.