Cursor: Inexpensive ones write code, expensive ones plan.

They got 100% test suite coverage for every configuration. Previously there were 70,000 merge conflicts. Now they are kept under 1,000 and we are able to generate 1,000 commits/sec instead of 1,000/hr.

With the “expensive planner + cheap workers” approach, the cheapest configuration ($1,339) which uses Opus 4.8 as the planner and cheap workers performs the same as GPT-5.5 ($10,565). This is possible because we have divided our tasks into two separate roles — planning and execution. Expensive models are responsible for planning while cheap models are responsible for executing. This allows us to reduce costs by 15x and decrease the total number of tokens used from 14.7 billion to 2.6 billion. This architecture enables us to scale up automated code generation.