A cheaper Cursor workflow: Opus for decisions, M3 for routine code
After about six months of using Cursor full time, Opus remained the only choice because its output was consistently strong. A week spent on major and several new caused the usage cost to rise sharply. Several cheaper models were tested, but most required so much manual correction that they saved little time.
handled straightforward work better, usually following the patterns already present in the code and needing only small fixes to names or unusual situations. The current routine uses Opus for roughly the first hour to understand the whole system, decide how should fit, and investigate difficult bugs; then handles and tests for the rest of the day.
Key points
- Cursor was used full time for about six months, initially with Opus alone.
- A major effort combined with several produced a painful usage cost.
- Most cheaper models needed enough manual editing to cancel out the time savings.
- usually matched existing coding patterns on straightforward tasks.
- Opus handles system-level decisions and difficult bugs, while handles and tests.