Subagents ignore Inherit from Parent / Auto Cost and always run as Auto Balance

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

When the parent agent runs with Auto (or when I configure subagent model pickup as Inherit from Parent or Auto Cost), spawned Explore/Task subagents always show and run as Auto Balance instead of inheriting the parent or using Auto Cost.

Evidence: in Agent chat, each subagent card shows the model label Auto Balance next to the task title (e.g. “Explore leanctx report/hooks · Auto Balance”, “Explore live-ab and docs · Auto Balance”), even though the intended pickup was Inherit from Parent or Auto Cost.

This looks related to known subagent model-routing issues (e.g. inherit falling back to Auto), but specifically here the wrong Auto variant is chosen: Balance instead of Cost / parent inherit.

Steps to Reproduce

  1. Open Cursor IDE Agent chat.
  2. Set the parent agent model/mode so subagent pickup should be Inherit from Parent, or set subagent model to Auto Cost.
  3. Ask the agent to spawn Explore/Task subagents (parallel exploration is enough).
  4. Observe the model label on each spawned subagent card in the chat UI.

Expected Behavior

  1. Inherit from Parent: subagents use the same model/mode as the parent agent.
  2. Auto Cost: subagents run under Auto Cost (not Auto Balance).
  3. UI label and usage dashboard match the configured pickup.

Operating System

MacOS

Version Information

Cursor: 3.22.12
Commit: 3a92974361033b2051526321308c2740fe5912c0
Arch: arm64
macOS: 26.7 (25G229)

For AI issues: which model did you use?

Parent: Auto (with subagent pickup set to Inherit from Parent and/or Auto Cost)
Subagents (actual): Auto Balance

For AI issues: add Request ID with privacy disabled

[Paste Request ID from the agent chat header / … menu after disabling privacy mode for that request]

Additional Information

Sometimes - I can sometimes use Cursor

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey @unknown_unknown, thanks for the report!

I checked your recent sessions, and the subagents are actually running on Auto Cost, the same as the parent chat. They’re billed as Auto on your usage, not as Auto Balance.

What you’re seeing is a labelling issue on the subagent card. The subagent inherits Auto from the parent without the Cost/Balance choice attached, so the card falls back to showing Auto’s default label, which is Balance. This is a known issue we’re already tracking, and I’ve added your report to it.

If you ever see a subagent that you think really ran on a different tier, reply with the request ID and I’ll check exactly what it ran on!

Thanks @Colin I will keep an eye.

Hey Colin, adding screenshot for your reference. Where we have initiated parent chat through Auto Cost, but sub agents got triggered on Claud-Opus-4.7-high-thinking model. Kindly help what could be the potential issue. Thanks

Hey @unknown_unknown, thanks for the screenshots, they made this easy to track down.

This one is different from the label issue I mentioned earlier. I checked that session, and the subagent really did run on Claude Opus 4.7 High. The parent chat (on Auto Cost) chose that model itself when it launched the subagent, instead of leaving it to inherit Auto. A subagent only inherits the parent’s model when the parent doesn’t name one, and the agent is only supposed to name one when you ask for it. We’re tracking this, and I’ve added your report to it.

A few things that help in the meantime:

  1. The “Explore Subagent Model” setting only applies to Explore subagents. The one in your screenshot was a general-purpose subagent, so that setting didn’t apply here.
  2. As a team admin, you can restrict which models your team can use from the team’s admin settings on the dashboard. Subagents can’t run a model that’s blocked for the team, so this is the reliable way to keep them off models like Opus.
  3. A rule like “When launching subagents, never set the model parameter” also helps steer the agent, but it isn’t a hard guarantee.

One question: did the design plan or your prompt mention Opus or a specific model anywhere? If it did, that would explain why the agent picked it, and it’s useful for us to know either way.