GMP - Dynamic Rules: leaderboard
Metric: F1 (x100) of the unsafe judgement averaged over GMP Task B's four context-specific rule sets (2,000 posts judged under each rule set's permissions and prohibitions, default-permit), zero-shot prompt with JSON output; higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.1 (Thinking) | 65 | #260 (DeepSeek V3.1) |
| 2 | Claude Sonnet 4 | 61.05 | #194 |
| 3 | Gemini 2.5 Flash | 60.58 | #237 |
| 4 | DeepSeek V3.1 (Non-reasoning) | 60.39 | #260 (DeepSeek V3.1) |
| 5 | Qwen 2.5 VL 72B Instruct | 58.52 | #364 |
| 6 | GLM-4.5 (Non-reasoning) | 57 | #265 (GLM-4.5) |
| 7 | GLM-4.5 (Thinking) | 55 | #265 (GLM-4.5) |
| 8 | GPT-4o Mini | 54.48 | #588 |
| 9 | Qwen 3 30B A3B 2507 Instruct | 54.31 | #464 |
| 10 | Kimi K2 | 53 | #236 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=gmp-dynamic-rules · How It Works · Data refreshed daily, snapshot 2026-10-11.