OpenAI GPT-6 Astra System Card - Adversarial User Simulations - Self-Harm: leaderboard
Table 7 (dynamic benchmarks with adversarial user simulations) of OpenAI's GPT-6 Astra System Card (Deployment Safety Hub, 2026-09-03): share of assistant messages that do not violate safety policy across evolving simulated self-harm conversations with adversarial users. Percent (the card prints 0-1), higher is better. GPT-6 Sol and GPT-6 Luna come from Table 30 of the card's section 11 (GPT-6 Sol and GPT-6 Luna, added 2026-09-22), which re-prints the other cells.
Metric: Safe message rate (%). Source: deploymentsafety.openai.com. Source: deploymentsafety.openai.com. Source: deploymentsafety.openai.com. Saturation forecast: Estimated already saturated. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6.1 Sol | 99.6 | #7 |
| 2 | GPT-6 Astra | 98.9 | #6 |
| 3 | GPT-5.4 (Thinking) | 97.5 | #83 (GPT-5.4) |
| 4 | GPT-6 Sol | 97.3 | #19 |
| 5 | GPT-6 Luna | 92.4 | #61 |
| 6 | GPT-6.1 Terra | 88 | #15 |
| 7 | GPT-5.5 | 86.8 | #33 |
| 8 | GPT-5.6 Sol | 85.6 | #22 |
No result here: #3 Claude Opus 5.5, #9 Gemini 4 Argon, #10 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=openai-gpt-6-astra-system-card-adversarial-user-simulations-self-harm · How It Works · Data refreshed daily, snapshot 2026-11-05.