PredAct-Bench: leaderboard
Final at-risk classification F1 on PredAct-CS, averaged across five target tool accuracies from 40% to 80%, with ten episodes per cell. The named model acts as the instructor and a fixed GPT-4o Mini assistant provides dialogue support.
Metric: Final F1 (%, PredAct-CS). Source: arxiv.org. Saturation forecast: Around 2031. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 63.4 | #99 |
| 2 | DeepSeek V4 Pro (Non-reasoning) | 60.7 | #102 (DeepSeek V4 Pro) |
| 3 | GPT-5.5 | 59.2 | #31 |
| 4 | Gemini 3.1 Pro (Preview) | 57 | #58 |
| 5 | Mistral Small 3.2 | 57 | #605 |
| 6 | Claude Opus 4.7 | 55.4 | #49 |
| 7 | GPT-4o Mini | 54.2 | #599 |
| 8 | DeepSeek V4 Flash (Non-reasoning) | 52.4 | #137 (DeepSeek V4 Flash) |
| 9 | Claude Haiku 4.5 | 51.9 | #280 |
| 10 | Qwen 3.5 35B A3B (Non-reasoning) | 49.8 | #256 (Qwen 3.5 35B A3B) |
| 11 | Ministral 3 14B | 49.2 | #650 |
| 12 | GPT-5.4 Mini | 44.4 | #212 |
| 13 | Qwen 3.5 9B (Non-reasoning) | 43.9 | #379 (Qwen 3.5 9B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=predact-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.