Vibe-Skills posts exact numbers behind its token savings

Vibe-Skills tagged version 4.1.0 on August 29, six weeks after v4.0.0. It posted +21.12 reward points and 29.6% fewer tokens against DeepSeekV4Flash-VE, measured on a benchmark called SkillsBench.

That specificity matters for teams evaluating token-reduction claims before they adopt one. Most repos on this beat post a vague percentage range and never say what they tested against, so there's nothing to rerun. Vibe-Skills names the model, the benchmark, and the exact delta instead.

The gains came from changes to skill-routing and workflow-orchestration between the two tags. The repo ships under an Apache-2.0 license and has 3,100 stars.

Nobody outside the repo has rerun SkillsBench against the new tag yet, so the 29.6% figure is checkable but not yet checked. Whether it survives an independent run decides if this is a working method or one team's internal score.