Hey Cursor Team,
I’ve noticed that the auto-router switches models mid-conversation (e.g., one message will go to Opus, the next to Grok). Previously, a best practice was to keep the same model for the entirety of the conversation for many reasons: prompt caching, system prompt, thinking tokens.
What’s your take there? Isn’t it resulting in worse quality overall? Do you have any tricks to make the mid switch possible?
-igor
Hey Igor, thanks for the request. This is expected behavior: in Auto mode, Cursor Router picks a model for each individual request, not one model for the whole chat. The idea is to choose the best model for each step.
If you need to stick to one model for the entire conversation (for prompt caching, system prompt consistency, and thinking behavior), the most reliable option right now is to manually select a specific model instead of Auto. Then it won’t switch mid-chat.
About routing: the chat context is still passed along no matter which model answers, so continuity isn’t lost. But yeah, the prefix cache is tied to a specific model, and when the model changes that benefit resets, so your intuition is right.
Also, there’s a related feature request about user-configurable routing rules. Feel free to check it out and add your vote or use case: # Feature Request: User-configurable model routing rules for the main agent
Thanks for a quick reply Dean; i guess I need to build a bit more experience with the auto router to to see if the mid-conversation switches bother me or not. The linked feature request is indeed interesting (assuming the switches are OK 
)