SAD - Predict Words: leaderboard
SAD INTROSPECT-PREDICT-WORDS asks a model to predict its own continuation of a random word list. Uses the authors' published adjusted score across task conditions; targets vary by model, so high accuracy need not imply an advantage over an external predictor. Uses the plain prompt condition.
Metric: Self-prediction score (%, higher is better). Source: github.com. Saturation forecast: Estimated already saturated. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 2 7B | 61.22 | #1498 |
| 2 | GPT-4 (0613) | 48.74 | #493 |
| 3 | Llama 2 13B | 48.67 | #1411 |
| 4 | Claude 3.5 Sonnet (20240620) | 48.41 | #308 |
| 5 | Llama 3 Chat 70B | 40.56 | #772 |
| 6 | GPT-4 Preview (0125) | 40.24 | #500 |
| 7 | Claude 3 Opus (20240229) | 37.96 | #428 |
| 8 | Claude 3 Sonnet (20240229) | 29.07 | #694 |
| 9 | GPT-4o | 26.84 | #357 |
| 10 | Claude 3 Haiku (20240307) | 25.77 | #780 |
| 11 | Claude Instant 1.2 | 23.61 | #1003 |
| 12 | Claude 2.1 | 22.18 | #833 |
| 13 | davinci-002 | 15.47 | #1400 |
| 14 | GPT-3.5 Turbo (0613) | 14.05 | #1106 |
| 15 | Llama 2 70B | 13.89 | #1039 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sad-predict-words · How It Works · Data refreshed daily, snapshot 2026-11-05.