CursorBench
CursorBench is a benchmark framework developed by Cursor for evaluating AI programming models. It systematically assesses the comprehensive capabilities of large language models on real-world programming tasks, covering scenarios such as code generation, code completion, and debugging. Through a standardized task set, it enables cross-model comparisons, providing developers and researchers with objective references for model performance.
Core Facts
CursorBench is a benchmark designed by the Cursor team specifically to evaluate the performance of AI programming models in real IDE environments
CursorBench simulates developers' daily tasks, including scenarios such as understanding cross-file dependency relationships, bug localization and fixing, and multi-module refactoring operations
Timeline (last 90 days)
Claude Opus 5.5 topped the CursorBench evaluation benchmark with a score of 57.8% (Max mode).
The official release disclosed only the single core data point of 57.8% on CursorBench, without detailing the gap relative to other models or performance on individual sub-tasks.
CursorBench is a vertical-scenario evaluation benchmark, distinct from general-purpose code generation benchmarks such as HumanEval and MBPP.
There is a gap between benchmark results and real-world experience, as a model's actual performance varies depending on the project's code style, technology stack, and task type.
In the CursorBench 4.0 benchmark, Grok 4.7 Extra High scored 46.3%, approaching Opus 5 Max's 46.6%, at a cost of only $6.01.
All Facts (4)
CursorBench is a benchmark designed by the Cursor team specifically to evaluate the performance of AI programming models in real IDE environments
80%VerifiedCursorBench simulates developers' daily tasks, including scenarios such as understanding cross-file dependency relationships, bug localization and fixing, and multi-module refactoring operations
75%UnverifiedCursorBench is a vertical-scenario evaluation benchmark, distinct from general-purpose code generation benchmarks such as HumanEval and MBPP.
50%UnverifiedThere is a gap between benchmark results and real-world experience, as a model's actual performance varies depending on the project's code style, technology stack, and task type.
50%