What is CursorBench?

CursorBench is Cursor’s own benchmark for coding agents, built by tracing committed code in its editor back to the request that produced it and turning each pair into a test case. Cursor’s technical report puts the median CursorBench task at around 180 changed lines, against 7 to 10 for SWE-bench, with prompts that are deliberately less specified, which is why the two benchmarks don’t rank the same model the same way.

Cursor hasn’t published how correctness is graded or how many tasks exist, and the tasks themselves are private, so nobody outside Cursor can check a result and Cursor also sells a model that’s scored on it. Terminal-Bench is one of the few independently maintained benchmarks that publishes both its method and its task set for the same kind of long, multi-file coding work, which makes it worth checking before trusting a CursorBench number on its own.