Vals AI SkillsBench: leaderboard
An agent completes a task with a Skill available, a folder of procedural instructions, scripts and resources. Tasks span eight domains, from software engineering and cybersecurity to natural science, finance and office work, each with a deterministic verifier. Vals AI runs three trials per task with the OpenHands CLI agent and reports the with-skills condition, averaged over trials then tasks; higher is better.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around May 2028. 35 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash (High) | 69.8 | #36 (DeepSeek V4.1 Flash) |
| 2 | Grok 4.5 (High) | 66.03 | #48 (Grok 4.5) |
| 3 | Gemini 3.7 Flash (High) | 65.89 | #23 (Gemini 3.7 Flash) |
| 4 | GPT-5.5 | 62.55 | #31 |
| 5 | GPT-5.5 (xHigh) | 62.21 | #31 (GPT-5.5) |
| 6 | Claude Fable 5.1 (Max) | 61.55 | #9 (Claude Fable 5.1) |
| 7 | GPT-5.6 Luna | 60.45 | #68 |
| 8 | Claude Opus 5 (Max) | 60.44 | #12 (Claude Opus 5) |
| 9 | Claude Opus 4.8 (Max) | 59.23 | #37 (Claude Opus 4.8) |
| 10 | Muse Spark 1.1 (xHigh) | 59.19 | #52 (Muse Spark 1.1) |
| 11 | GPT-5.6 Terra | 58.9 | #41 |
| 12 | Gemini 3.8 Flash (High) | 57.98 | #21 (Gemini 3.8 Flash) |
| 13 | Grok 4.6 (High) | 55.77 | #26 (Grok 4.6) |
| 14 | Qwen 3.7 Plus | 54.3 | #92 |
| 15 | GPT-5.6 Sol (Max) | 54.1 | #19 (GPT-5.6 Sol) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-skillsbench · How It Works · Data refreshed daily, snapshot 2026-10-20.