SocietyBench - Calibration: leaderboard
The calibration axis of SocietyBench (2026): models give probabilities for whether events happen after each cutoff on anonymized timelines of five real social events, scored as 100 times one minus the weighted mean absolute error (0-100, a uniform 50 percent answer scores 50). Higher is better.
Metric: Calibration score (0-100): 100 x (1 - weighted mean absolute error of event probabilities), weighted by resolution distance and question window; five real social events (Wuhan Library, US-Iran, TikTok, SMCI, Trump Tariff) anonymized into counterfactual worlds, 25 prediction points each, averaged over the Chinese and English editions; temperature 0; trivial anchor 50. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 75.3 | #49 |
| 2 | Gemini 3.5 Flash | 73.9 | #101 |
| 3 | Seed 2.0 Pro | 64.3 | #211 |
| 4 | Kimi K2.5 | 63.3 | #256 |
| 5 | DeepSeek V4 Pro | 62.5 | #188 |
| 6 | Claude Opus 4.8 | 62.5 | #56 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=societybench-calibration · How It Works · Data refreshed daily, snapshot 2026-11-13.