OpenAI GPT-5.6 System Card - Adversarial User Simulations - Self-Harm: leaderboard
Table 7 (dynamic benchmarks with adversarial user simulations) of OpenAI's GPT-5.6 System Card (Deployment Safety Hub, 2026-07-09): share of assistant messages that are not unsafe across evolving simulated self-harm conversations with adversarial users. Percent (the card prints 0-1), higher is better.
Metric: not_unsafe message rate (%). Source: deploymentsafety.openai.com. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Thinking) | 97.7 | #83 (GPT-5.4) |
| 2 | GPT-5.2 (Thinking) | 95.5 | #115 (GPT-5.2) |
| 3 | GPT-5.6 Terra | 94.7 | #40 |
| 4 | GPT-5.6 Luna | 90.5 | #79 |
| 5 | GPT-5.1 (Thinking) | 90.4 | #145 (GPT-5.1) |
| 6 | GPT-5.5 | 86.8 | #33 |
| 7 | GPT-5.6 Sol | 85.6 | #22 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openai-gpt-5-6-system-card-adversarial-user-simulations-self-harm · How It Works · Data refreshed daily, snapshot 2026-11-05.