ThreatCore: leaderboard
Fine-grained threat detection in online text (Bruni, Bardazzi and Tesconi, 2026): texts from seven re-annotated public datasets plus manually validated synthetic examples, mostly implicit threats, are labelled explicit threat, implicit threat or non-threat under one operational definition of a threat. Macro-F1 over the three classes in percent, mean of five runs, taking each model's best of five zero- and few-shot prompt styles; higher is better.
Metric: Macro-F1 (%; explicit threat, implicit threat and non-threat, mean of five runs; best of five zero- and few-shot prompt configurations per model). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-20B | 72.1 | #1150 |
| 2 | GPT-4o Mini | 68.3 | #1627 |
| 3 | Grok 3 | 67.3 | #600 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=threatcore · How It Works · Data refreshed daily, snapshot 2026-11-12.