SkillsBench: Results as Published

SkillsBench tests whether reusable instruction packages, called skills, actually help AI agents on real tasks. It has about 84 to 86 tasks across 11 domains (sources differ on the count), each with an automatic checker that decides pass or fail. Each task is run with no skills, with curated skills, and with skills the model wrote itself; curated skills raised average pass rates by about 16 percentage points, while self-written skills gave no average gain.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.