SAD - Mini: leaderboard
SAD-mini uses portable multiple-choice tasks on model facts, influence and deployment stages. It excludes the self-prediction and self-recognition tasks, so its score does not measure behavioral self-prediction. Uses the plain prompt condition.
Metric: Score (%, higher is better). Source: situational-awareness-dataset.org. Saturation forecast: Estimated already saturated. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 Preview (2024-09-12) | 74.94 | #214 |
| 2 | O1 Mini (2024-09-12) | 70.97 | #412 |
| 3 | Claude 3 Opus | 66.74 | #565 |
| 4 | Claude 3.5 Sonnet | 66.49 | #362 |
| 5 | Claude 3 Sonnet | 61 | #774 |
| 6 | Llama 3 Chat 70B | 60.64 | #772 |
| 7 | GPT-4 (0613) | 60.62 | #493 |
| 8 | GPT-4o | 59.44 | #357 |
| 9 | GPT-4 Preview (0125) | 57.66 | #500 |
| 10 | Claude 2.1 | 57.34 | #833 |
| 11 | Claude Instant 1.2 | 56 | #1003 |
| 12 | Claude 3 Haiku | 54.17 | #801 |
| 13 | Llama 2 70B Chat | 50.9 | #1275 |
| 14 | Llama 2 70B | 49.59 | #1039 |
| 15 | Llama 2 13B Chat | 48.92 | #1414 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sad-mini · How It Works · Data refreshed daily, snapshot 2026-11-05.