SkillsBench
Whether an agent uses packaged tools and skills effectively, scored both with and without the skill library available.
Agentic toolingVals AI35 models run by Vals AIUpdated 27 September 2026
How it is run
Reports two numbers per model, which makes the difference between them the interesting quantity.
What it cannot tell you
A model that scores well with skills and poorly without them is measuring the skill library as much as the model.
- Leader
- DeepSeek V4.1 Flash69.80% · Vals AI's own list
- Top-ten spread
- 10.61%Leader minus tenth place, or last place in smaller fields.
- Median
- 52.74%of the 35 rows published
- No published result
- 26of 46 models on this site
Independent leaderboard · Vals AI
Top 35 of 35 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | DeepSeek V4.1 Flash | 69.80% |
| 2 | Grok 4.5 | 66.03% |
| 3 | Gemini 3.7 Flash | 65.89% |
| 4 | Gpt 5.5 Codex | 62.56% |
| 5 | GPT-5.5 | 62.21% |
| 6 | Claude Fable 5.1 | 61.55% |
| 7 | GPT-5.6 Luna | 60.45% |
| 8 | Claude Opus 5 | 60.44% |
| 9 | Claude Opus 4.8 | 59.23% |
| 10 | Muse Spark 1 1 | 59.19% |
| 11 | GPT-5.6 Terra | 58.90% |
| 12 | Gemini 3.8 Flash | 57.99% |
| 13 | Grok 4.6 | 55.77% |
| 14 | Qwen3.7 Plus | 54.30% |
| 15 | GPT-5.6 Sol | 54.10% |
| 16 | DeepSeek V4 Pro | 53.83% |
| 17 | Muse Spark 1.2 | 53.04% |
| 18 | Gemini 3.5 Flash | 52.74% |
| 19 | Gpt 5.4 2026 03 05 | 51.71% |
| 20 | MiniMax M3 | 51.50% |
| 21 | Deepseek V4 Pro | 51.27% |
| 22 | Deepseek V4 Flash 0731 | 50.67% |
| 23 | Kimi K2.7 Code | 50.04% |
| 24 | Claude Sonnet 4 6 | 49.05% |
| 25 | GLM 5.3 | 47.51% |
| 26 | Claude Sonnet 5 | 46.48% |
| 27 | Glm 5.2 | 45.08% |
| 28 | Qwen3.8 Max | 42.01% |
| 29 | Grok 4.3 | 40.64% |
| 30 | GLM 5.3 Flash | 40.16% |
| 31 | Qwen3.8 27b | 38.11% |
| 32 | Inkling Small | 33.58% |
| 33 | Ling 3.0 Flash 2607 | 27.95% |
| 34 | Inkling | 26.09% |
| 35 | Mercury 2.5 | 18.06% |
No published result tracked
26 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.
Step 5 PreviewGPT-6 LunaQwen3.8 2.4T A95BQwen3.8-Flash-NextMiMo-V2.6-FlashQwen3.8 Max (0902)Claude Fable 5Claude Opus 4.7GPT-5.4DeepSeek V4.1 Flash 1017GLM-LegalClaude Haiku 5.5BeamGemini 4 ArgonClaude Sonnet 5.5Claude Opus 5.5GPT-6 AstraGPT-6 SolGPT-6.1 SolGemini 3.1 ProGrok 4.7Kimi K3Muse Spark 1.3 MaxHunyuan Hy4 PreviewMiMo V2.6 ProMistral Medium 3.5
Read this with
Benchmarks in the same category, which usually disagree with this one on purpose.
Sources
Every page used for this benchmark, read on 17 October 2026.
- Benchmark site or the page that publishes itwww.skillsbench.ai/
- Independent resultswww.vals.ai/benchmarks/skillsbench