Cursor burned $5 per request on cache reads. I found a bug, reported it, and then discovered something better. Aider gave me control I never had with any IDE.
Hi, my name is Tom Smykowski, I'm a staff full-stack engineer. I build and scale SaaS platforms to millions of users, working end-to-end from system architecture to frontend to mobile. On this blog I share practical insights on AI-powered development and cost optimization.
When I design SaaS architectures or build AI-powered workflows, I track every dollar. Last month, my Cursor bills made me stop and investigate. What I found changed how I think about AI coding tools entirely.
I've lately found that Cursor burns a lot of tokens on cache reads. After some research, I believe there's a bug causing it. I've reported it to the Cursor team, and they're investigating. But this discovery motivated me to search for better options for part of my workflows.
Even if the bug gets fixed, it's no mystery that Cursor, Claude Code, and Windsurf are optimized for development in ways that aren't perfectly tailored to all needs. They want to offer something useful for professional engineers while also appealing to the general public. The result is minimal systems where you don't have much control over context, cache, and other aspects.
The special category in my workflow that brought my attention lately are what I call AI scripts. These combine deterministic and non-deterministic processes. They come in various shapes, but the one I was interested in optimizing was a script governed by AI that reads files, calls other scripts, and uses them to provide output.
In the process, I was able to lower the cost of execution from $5 to $0.17. That's the difference between something that doesn't scale and something that does.