Cursor's auto mode burned 35M tokens on a sub-agent for barely any work
While using Cursor's 'auto' model-selection feature, the system spun up a sub-agent (a secondary AI instance spawned to handle a sub-task) on its own, and that sub-agent consumed 35 million tokens on a single prompt. Almost no actual work got done despite that.
The same person had a similar experience two days earlier: while using Cursor-Grok-4.5, the system silently swapped the model to Thinking High without being asked. This time, using auto mode, the same kind of unexpected sub- happened again.
This single issue cost $55 over two days. The person asked Cursor's whether this is intended behavior and how to disable it, and asked other users whether they've hit the same problem and how they avoid it.
Key points
- While using Cursor's auto mode, the system spawned a sub-agent on its own that consumed 35 million tokens in a single prompt
- Very little actual work got done despite the massive
- A similar incident happened two days earlier: the model silently switched from Cursor-Grok-4.5 to Thinking High
- This issue alone cost $55 over two days
- It's unclear whether this is intended behavior or if there's a way to disable it