SAD - Anti-Imitation: leaderboard
SAD anti-imitation tasks test whether a model can depart from a likely continuation and deliberately control its output distribution. 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) | 29 | #214 |
| 2 | Claude 3.5 Sonnet | 27.09 | #362 |
| 3 | Llama 2 13B | 19.89 | #1411 |
| 4 | davinci-002 | 15.39 | #1400 |
| 5 | Llama 2 70B | 13.11 | #1039 |
| 6 | Claude 3 Opus | 12.14 | #565 |
| 7 | GPT-4o | 11.18 | #357 |
| 8 | Claude 2.1 | 10.25 | #833 |
| 9 | GPT-4 Preview (0125) | 10.04 | #500 |
| 10 | Llama 2 7B | 9.32 | #1498 |
| 11 | GPT-3.5 Turbo (0613) | 6.73 | #1106 |
| 12 | GPT-4 (0613) | 6.36 | #493 |
| 13 | O1 Mini (2024-09-12) | 5 | #412 |
| 14 | Claude 3 Sonnet | 4.34 | #774 |
| 15 | Claude 3 Haiku | 3 | #801 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sad-anti-imitation · How It Works · Data refreshed daily, snapshot 2026-11-05.