Dev.to
5/13/2026

I asked Cursor to rename a function. It sent 8,400 tokens. I checked.
Short summary
An engineer discovered Cursor was sending 6,500+ unnecessary tokens for simple tasks, inflating their API bill 50% monthly. They built a 200-line routing layer (Haiku for simple, Sonnet for code, Opus for planning) and reduced costs 41% while increasing call frequency. As users discover they can route their own model calls, AI tool wrappers face pricing pressure—seats will likely drop from $20 to $5-10/month.
- •Cursor was adding 6,500 excess context tokens to simple refactoring tasks due to conservative routing defaults
- •Built custom 200-line TypeScript router that cut API costs 41% by selecting models based on prompt intent classification
- •AI tool wrappers now shipping model dropdowns because users increasingly realize they can route calls themselves
Generated with AI, which can make mistakes.
Is this a good recommendation for you?



