CursorBench is Cursor’s internal evaluation suite for coding agents, run on ambiguous, multi-file tasks taken from real Cursor IDE sessions. Tasks span codebase understanding, bug finding, editing, refactoring, planning, and code review. Each model is evaluated at several reasoning-effort levels, and the leaderboard reports a correctness score along with the average cost, token usage, and number of agent steps per task.

We source results from the public CursorBench leaderboard. Our chart reports the correctness score, where higher is better. Most models appear once per reasoning-effort level (Low, Medium, High, Extra High, Max), and we keep that level along with per-task cost, token usage, and step count in the data export.

CursorBench tasks are collected from real Cursor sessions and are designed to be ambiguous and span multiple files, so they test agentic, repo-aware coding rather than isolated code completion. Because each reasoning level trades off score against cost and latency, we keep all levels as separate entries rather than collapsing to a single number per model.