Dear Cursor Team,
I’ve been using Cursor with my entire team for the past 3 months, and I’d like to share a suggestion based on our experience.
As full-time developers, we found that the $20/month plan offers fast requests, but they’re consumed within a day of normal usage. For the remaining 29 days, we’re left with slow requests. Recently, we noticed that even when selecting Sonnet 3.7, slow requests are now routed through the “auto” model instead. Unfortunately, this often results in unusable code—sometimes even destructive changes like auto-installing packages or committing broken code.
It seems the root problem might be this:
When the agent generates poor code, it takes multiple attempts to fix => more tokens are consumed.
When the agent generates good code from the start => fewer tokens are needed.
Ironically, this means users are penalized (with higher token usage and worse performance) when the AI agent is less effective. That doesn’t feel fair. We hope you’ll consider improving the token calculation model—or better yet, allow users to add their own API keys to maintain fast request speeds (or at least access reliable models like Sonnet even for slow requests).
Currently, the “auto” model isn’t helping us, and we’re testing other tools because of this limitation. We hope to continue using Cursor, and would love to see improvements in this area.
Thank you for your great work and consideration.
Best regards,
Hodela