Cursorbench

CursorBench is Cursor's own test of coding agents, built from real tasks that Cursor users gave the agent in their editor. Cursor traces committed code back to the request that produced it, and turns those requests into test cases. Version 4.0 came out in September 2026 and added longer tasks: refactors, investigations, working out what a vague request actually meant, and sticking to a design.

The leaderboard reports a correctness score plus the cost, tokens, and number of steps each model needed, and it scores each model at several reasoning-effort settings. Cursor has not published how correctness is graded or how many tasks there are, and the tasks themselves are private. The top score at the time of writing is about 52%.

Its strength is that the tasks are the messy, multi-file, underspecified requests real developers type, drawn from codebases models haven't trained on. Its weakness is that only Cursor can run it, so nobody can check a result, and Cursor also sells a model that scores well on it. Third-party benchmarks have ranked that same model much lower. Treat CursorBench as a signal about how models behave inside Cursor, not as a neutral ranking.

Cursor refreshes the task set regularly, so scores across versions don't compare, and 4.0's longer tasks pulled every score down from the previous version.