FinSkillBench (Self-Generated Skills): leaderboard
FinSkillBench (Deep Insight Labs, 2026) tests whether language-model agents use financial domain skills to solve investment-management tasks, scoring 12 subtasks against hidden ground truth with task-specific verifiers. The score is the unweighted mean subtask score on a native 0 to 1 scale, higher is better, where the agent writes and reuses its own skill documents within each episode.
Metric: Task score (0-1, native; the agent writes and reuses its own skill documents within each episode; over FinSkillBench's 12 investment-management subtasks (portfolio construction, risk management, fundamental analysis), unweighted mean, hidden ground truth and task-specific verifiers). Source: arxiv.org. Saturation forecast: Around December 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0.49 | #158 |
| 2 | DeepSeek V3.2 | 0.45 | #403 |
| 3 | Gemini 3.1 Flash Lite | 0.43 | #363 |
| 4 | Gemini 2.5 Pro | 0.41 | #262 |
| 5 | GPT-5.4 | 0.41 | #121 |
| 6 | GPT-4.1 | 0.4 | #492 |
| 7 | GLM-5.1 | 0.25 | #175 |
| 8 | Grok 4 | 0.14 | #306 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=finskillbench-self-generated-skills · How It Works · Data refreshed daily, snapshot 2026-11-12.