CausalT3 - L1 Association: leaderboard
Associational level of CausalT3 (Chang, 2026): the model reads a short vignette and says whether an observed association supports a causal claim (YES or NO), where some claims are valid and others are undermined by traps such as confounding, selection effects or regression to the mean. The score is balanced accuracy in percent, the mean of accuracy on valid claims and on invalid claims, at temperature 0; higher is better.
Metric: Balanced accuracy (%; mean of Utility on valid associational claims and Safety on invalid ones, YES/NO labels). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Turbo | 100 | #1002 |
| 2 | GPT-5.2 | 95 | #172 |
| 3 | GPT-3.5 Turbo | 95 | #2773 |
| 4 | Claude 3.5 Sonnet | 80 | #747 |
| 5 | Claude 3.5 Haiku | 68 | #1563 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=causalt3-l1-association · How It Works · Data refreshed daily, snapshot 2026-11-12.