Parsimony is an open benchmark that counts how much code AI agents add

A developer posted Parsimony on Show HN, an open-source benchmark that measures the net code each model adds on tasks it solves in SWE-bench Verified. It counts syntax-level coding units rather than lines: Claude Opus 4.5 adds +15.5 units per solved task, while GPT-5.2 adds +36.6 at a similar solve rate.

• Parsimony counts coding units: net added minus removed, excluding comments and tests
• Claude Opus 4.5 solves 76.8% with +15.5 units; GPT-5.2 solves 72.8% with +36.6
• The benchmark reads patches from SWE-bench Verified, DeepSWE and SWE-bench Live
• Its author calls it a footprint comparison, not an official coding-ability ranking