SAD - Lite: leaderboard
SAD-lite excludes the model-facts, multi-hop identity and token-counting tasks that need extra model-specific setup. Published subset score for situational awareness and self-knowledge. 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) | 58.78 | #214 |
| 2 | Claude 3.5 Sonnet | 55.01 | #362 |
| 3 | O1 Mini (2024-09-12) | 52.07 | #412 |
| 4 | Claude 3 Opus | 51.02 | #565 |
| 5 | GPT-4o | 47.73 | #357 |
| 6 | Claude 3 Sonnet | 46.42 | #774 |
| 7 | Llama 3 Chat 70B | 45.26 | #772 |
| 8 | Claude 2.1 | 44.82 | #833 |
| 9 | GPT-4 Preview (0125) | 44.65 | #500 |
| 10 | GPT-4 (0613) | 44.46 | #493 |
| 11 | Claude Instant 1.2 | 42.39 | #1003 |
| 12 | Claude 3 Haiku | 41.04 | #801 |
| 13 | Llama 2 70B Chat | 37.48 | #1275 |
| 14 | GPT-3.5 Turbo (0613) | 36.89 | #1106 |
| 15 | Llama 2 7B | 36.49 | #1498 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sad-lite · How It Works · Data refreshed daily, snapshot 2026-11-05.