Does a high CursorBench score mean an agent will work in production?

Not really: Cursor builds the tasks, runs the grading harness, and sells the model being scored, all under one roof. Cursor’s own Composer 2 report shows the asymmetry directly: for public benchmarks like Terminal-Bench, it publishes its own harness’s score next to the officially reported one so a reader can compare the two, but no second number exists for CursorBench, for any model, because no outside party can run it.

The same report also specifies what “production” means here: the agent runs exactly as it would inside Cursor’s own production environment, which describes how Cursor operates the model, not how the code that model writes holds up once it ships. A model that tops CursorBench is still one you’d want to check against your own production traces rather than a vendor’s own leaderboard.