Cursor publishes CursorBench 3.1 coding-agent evaluation

Category: Agentic coding
Status: discussion
Sources: CursorBench 3.1, HN discussion
Summary: Cursor published CursorBench 3.1, a vendor evaluation of coding agents on ambiguous multi-file tasks drawn from real Cursor sessions, covering codebase understanding, bugfinding, planning, and code review. The page ranks 36 model configurations and describes its per-task cost calculation from published per-token pricing, but does not disclose task construction or grading rubrics.
Why it matters: CursorBench is a signal on how frontier models perform inside one widely used agent harness, though as a vendor benchmark without a published rubric it is not an independent result.
Follow-up: Track whether Cursor publishes task construction and grading methodology or an independent replication appears.