Show HN: Parsimony – an open-source benchmark for how much code AI models add

Parsimony is a new open-source benchmark that ranks AI models by mean net coding units added (added minus removed) on SWE-bench Verified tasks they solved, rather than just solve rate or cost. It counts syntax elements instead of lines, excludes tests, docs and generated files, and only measures Python. Developers choosing models for coding agents can use it to see which models produce leaner patches, though the author notes more code often correlates with more tasks solved.

Read the full story on parsimony.lasu.dev

Source

parsimony.lasu.dev

All items on the wire