Parsimony: open-source benchmark ranks AI models by net code added

Parsimony, an open-source benchmark released by lasu, measures how much code AI models add to patches, counting net coding units with deleted code subtracted rather than raw lines.

It ranks models by mean net units added on SWE-bench Verified tasks, lowest first, and includes failed attempts in the mean. The creator says existing benchmarks measure solve rate and cost but do not penalize excess code, which makes a codebase harder to work with long term. Only Python is measured, and the benchmark does not rerun tests.