SAD - Introspection: leaderboard
SAD introspection combines token counting, prediction of the model's own word continuations and checking whether its outputs obey rules. Use Predict Words for the separate behavioral self-prediction task. 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 | Claude 3.5 Sonnet | 39.1 | #362 |
| 2 | GPT-4 (0613) | 38.32 | #493 |
| 3 | Llama 2 7B | 37.95 | #1498 |
| 4 | Llama 2 13B | 35.02 | #1411 |
| 5 | Llama 3 Chat 70B | 34.81 | #772 |
| 6 | GPT-4 Preview (0125) | 34.09 | #500 |
| 7 | Claude 3 Opus | 33.64 | #565 |
| 8 | Claude Instant 1.2 | 30.27 | #1003 |
| 9 | Claude 3 Sonnet | 30.18 | #774 |
| 10 | GPT-4o | 29.64 | #357 |
| 11 | O1 Preview (2024-09-12) | 27.93 | #214 |
| 12 | Claude 3 Haiku | 27.6 | #801 |
| 13 | O1 Mini (2024-09-12) | 26.58 | #412 |
| 14 | Claude 2.1 | 26.01 | #833 |
| 15 | davinci-002 | 23.98 | #1400 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sad-introspection · How It Works · Data refreshed daily, snapshot 2026-11-05.