Frontier AI Risk Monitor - APE: leaderboard
In the Attempt to Persuade Eval the model is asked to persuade a simulated user toward harmful or controversial claims; the platform reverses the persuasion-attempt rate, so the number is the share of 200 conversations where the model did not try. It measures willingness to persuade, not persuasive skill (MakeMePay and PMIYC measure skill).
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 54 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 72.85 | #5 |
| 2 | Grok 4.7 | 70.85 | #44 |
| 3 | Claude Opus 4.5 (Thinking) | 69 | #156 (Claude Opus 4.5) |
| 4 | GPT-6 Astra | 67.65 | #6 |
| 5 | GPT-5.4 (High) | 66.35 | #123 (GPT-5.4) |
| 6 | GPT-5.6 Sol | 62.77 | #30 |
| 7 | Claude Opus 4.6 (Thinking) | 62.29 | #102 (Claude Opus 4.6) |
| 8 | GLM-5 | 61.92 | #260 |
| 9 | GPT-5.5 | 60.94 | #49 |
| 10 | GPT-5.1 (High) | 59.04 | #229 (GPT-5.1) |
| 11 | Kimi K2.6 | 58.31 | #178 |
| 12 | Qwen 3.5 Plus (Thinking) | 57.73 | #232 (Qwen 3.5 Plus) |
| 13 | MiniMax-M2 | 54.1 | #674 |
| 14 | MiniMax-M2.7 | 51.46 | #518 |
| 15 | GPT-5 (High) | 51.06 | #167 (GPT-5) |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-ape · How It Works · Data refreshed daily, snapshot 2026-11-13.