SkillsBench: Benchmarking How Well Skills Work Across Diverse Tasks
The first evaluation framework that measures how skills work — and the first dataset that measures how well models use skills on expert-curated tasks across high-GDP-value domains.
Agent performance
Resolution rate against mean agent wall-clock per task (log scale, faster to the right). The no-skills counterparts are ghosted for context.
v1.1 leaderboard
Resolution rate with the full skills bundle loaded. Δ is the change against the same harness with no skills available.
| # | Model | Harness | Resolution | Δ vs no-skills | Wall-clock | |
|---|---|---|---|---|---|---|
| 01 | GPT-5.5 | Codex | 63.4% | +18.2 | 6m 20s | |
| 02 | Opus 4.8 | Claude Code | 61.9% | +21.4 | 5m 04s | |
| 03 | Gemini 3.1 Pro | Gemini CLI | 57.2% | +14.9 | 9m 11s | |
| 04 | Opus 4.7 | Claude Code | 55.8% | +19.6 | 5m 42s | |
| 05 | GLM 5.1 | OpenHands | 52.1% | +16.0 | 12m 30s | |
| 06 | Kimi K2.6 | OpenHands | 48.7% | +12.7 | 14m 02s | |
| 07 | MiniMax M3 | OpenHands | 44.3% | +9.8 | 16m 44s | |
| 08 | Sonnet 4.6 | Claude Code | 43.0% | +17.1 | 4m 12s | |
| 09 | DeepSeek V4 Pro | OpenHands | 41.6% | +11.3 | 18m 09s | |
| 10 | Haiku 4.5 | Claude Code | 31.2% | +13.8 | 2m 51s |
How SkillsBench works
Three abstraction layers, mirroring how traditional computing systems are structured.
Skills layer
Domain-specific capabilities and workflows that extend agent function. Like applications on an OS, skills carry specialised knowledge and tools for particular tasks.
Agent harness layer
The execution environment that orchestrates agents, manages tool access and handles I/O. Analogous to an operating system mediating between applications and hardware.
Models layer
The foundation models that power reasoning and generation. Like CPUs, they provide the raw computational capability every layer above depends on.