Show HN: Parsimony Benchmarks How Much Code AI Models Add

Summary

Parsimony is an open-source benchmark designed to measure how much code AI models add while solving software tasks, addressing a gap in benchmarks that track solve rate and token cost but do not penalize excessive code. It uses patches published by DeepSWE, SWE-bench Verified, and SWE-bench Live, then counts syntax-level “coding units” such as statements, calls, names, and literals. The benchmark does not use lines of code, so formatting or compressed code does not improve the measurement. Deleted code is subtracted to produce a net-unit score. GPT-6 and Claude 5.5 had no public patches available, so the creator ran those models on 3–10 tasks using personal subscriptions; those results are shown as faded circles and are not ranked. The author is seeking feedback and public patch exports for newer models. Parsimony is not affiliated with the existing benchmarks.