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

Parsimony is an open-source benchmark that measures how much code AI models add when submitting patches, focusing on code growth rather than just pass/fail rates or compute cost. It computes net "coding units" - syntax elements such as statements, calls, comparisons, names, and literals - by parsing before-and-after versions of changed Python files and subtracting removed units from added ones. Tests, docs and generated files are excluded, patches are not executed, and pass/fail status is taken from SWE-bench’s published results. Models are ranked by mean net units added across solved tasks (including failed attempts in that metric), with an interactive leaderboard and charts driven by JSON data hosted on GitHub.

Findings and caveats are explicit: smaller output is not always better because unit counts measure size, not readability or design, and models that produce more code tend to solve more tasks (Spearman 0.33-0.60 on boards with 26+ models). All models run through the same mini-SWE-agent, which still influences code volume. Extra features include an Artificial Analysis Intelligence Index (informational only), author-run unverified "faded" model runs, an example refactor to illustrate reduced units, and tooling to import other benchmark results so teams can compare code-size impacts alongside classical evaluation metrics.