Parsimony Benchmark Tracks How Much Code AI Models Add
Problem-solving scores and cost have carried the comparison between coding models, and Parsimony makes code growth a comparable number, though its developer says code volume alone cannot measure readability or design quality.
Reporting from 1 source: GIGAZINE.
An open-source benchmark called Parsimony measures the code AI adds minus the code it deletes when it modifies a program. It analyzes existing benchmark modification data using a unit that counts syntactic elements rather than lines, so line breaks and comments do not count. In its published comparison, Claude Opus 5.5(high) adds about 1.1 times as much code as other models, while GPT-6.1 Sol(high)/Codex adds about 0.7 times.
The developer says the benchmark was built to measure bloat, since code that solves a problem can still add unnecessary processing that makes review and long-term maintenance harder. Test code, documentation, and auto-generated files are excluded from the count. Because deletions subtract from additions, a rewrite that simplifies existing processing can score a negative value, such as replacing a summing loop with Python's standard "sum" function, which the developer cites as a net reduction of 10 units.
The source code and measurement data are on GitHub and the results can be reproduced from that data. The developer is asking for modification data from new models and for opinions on evaluation methods.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.