SkillBench

Which AI actually does professional work best?

There's HumanEval for code and MMLU for knowledge — nothing measures whether a model can write a PRD your team can execute, a postmortem that finds the real root cause, or a board update that survives the room. SkillBench is that benchmark: a fixed set of realistic professional tasks, a published rubric, and a pinned LLM judge — fully reproducible from this repo.

Loading results…

🏆 Model rankings

English task set · 中文任务集 (Chinese task set: DeepSeek, Qwen, Kimi, GLM and Doubao on Chinese workplace, life-admin, exam and content tasks)

Loading…

🥊 Agent Arena

Community-run results for any model or agent behind an OpenAI-compatible endpoint, on the same tasks and rubric as the rankings above. Every row links to its public run or its saved outputs, and the score is recomputed from the per-task numbers before it is ranked. A row graded by a judge other than the pinned one is marked, because its score is not strictly comparable.

Loading…

🔬 How it's scored

Each model runs every task two ways — bare (task only) and skilled (task + the matching SKILL.md). The judge scores each output 1–5 on the same fixed rubric the library uses — structure, completeness, usefulness, grounding — with the judge model + version pinned per release. The headline SkillBench score is the mean skilled score; skill lift is skilled minus bare.

📖 Methodology & task set 🎯 Skill leaderboard (scores skills, not models) ℹ️ About the rubric

▶ Run it yourself

SkillBench is open and reproducible — add a model to the board by running the harness and committing the result. Estimate cost first with --dry-run.

# Score one model (Anthropic / OpenAI / Gemini via env keys)
ANTHROPIC_API_KEY=… node skillbench/run-skillbench.mjs --models claude-sonnet-4-6

# Compare across providers
ANTHROPIC_API_KEY=… OPENAI_API_KEY=… GEMINI_API_KEY=… \
  node skillbench/run-skillbench.mjs --models claude-opus-4-8,gpt-4o,gemini-2.0-flash