Benchmarks & Deep Dives · Free
CursorBench
CursorBench is Cursor's internal benchmark for AI coding agents, built from ambiguous, multi-file tasks drawn from real Cursor sessions rather than from public GitHub issues or short coding puzzles. Cursor publishes the current version, CursorBench 4.0, as a public leaderboard that ranks model and effort configurations by score, and lists each one's cost per task, tokens per task and steps per task alongside the score. That combination makes it one of the few public coding evaluations that shows the price of a result, not just the result itself.
What it measures: whether an agent running inside Cursor's own harness can complete realistic developer requests that span several files and leave room for interpretation. Because the tasks come from Cursor's product, the benchmark reflects how models behave in an IDE agent loop with tool calls, file edits and verification steps. Results are run by Cursor, not by model vendors, though the task set itself is not public.
How to read it: a higher score is better, but cost per task matters just as much for production budgets. The same model often appears several times at different effort settings, and the leaderboard regularly shows that the highest effort is not always the most efficient. In November 2026, for example, Gemini 4 Carbon at high effort topped CursorBench 4.0 at 60.4% at about $9.22 per task, while Claude Fable 5.2 at max effort scored 60.0% at about $32.79 per task. Cursor's page does not state which price list it used for models with temporary introductory pricing, so check vendor price changes before relying on cost columns.
Limitations: the tasks are private, so results cannot be reproduced independently; the scores depend on Cursor's harness, prompts and tool set; and cost figures move whenever vendors change their prices. As with SWE-bench and Terminal-Bench, treat CursorBench as one signal among several and confirm with an evaluation on your own repositories.
Who it is for: engineering leaders choosing a default model for coding agents, teams comparing effort settings to control spend, and anyone tracking cost per successful task rather than cost per token.
Pricing type: free to view.
Alternatives: SWE-bench, Terminal-Bench, Zapier's AutomationBench and the Artificial Analysis Coding Agent Index.
Official URL: https://cursor.com/cursorbench
