ConfidenceBench - Brier Score: leaderboard
Verbalized-confidence calibration on 200 multiple-choice questions in four categories, pooled over three runs: the Brier score of each stated confidence against whether the answer was right. Refusals are excluded rather than scored wrong. Lower is better; a calibrated-random baseline scores 0.1875. Source: arxiv.org/abs/2607.20526.
Metric: Brier Score (0-1). Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.1 | #64 |
| 2 | Claude Opus 4.6 | 0.1 | #67 |
| 3 | GPT-5 | 0.12 | #105 |
| 4 | GPT-5 (High) | 0.12 | #105 (GPT-5) |
| 5 | GPT-5 Mini | 0.13 | #193 |
| 6 | GPT-5 Nano | 0.13 | #439 |
| 7 | GPT-5 (Low) | 0.14 | #105 (GPT-5) |
| 8 | Claude Sonnet 4.6 | 0.14 | #94 |
| 9 | Claude Opus 4.5 | 0.15 | #87 |
| 10 | Claude Sonnet 4.5 | 0.18 | #148 |
| 11 | Claude Opus 4.1 | 0.19 | #146 |
| 12 | Gemini 2.5 Pro | 0.21 | #159 |
| 13 | Gemini 2.5 Flash | 0.23 | #255 |
| 14 | Claude Sonnet 4 | 0.25 | #207 |
| 15 | Gemini 3.1 Flash Lite | 0.37 | #200 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=confidencebench-brier-score · How It Works · Data refreshed daily, snapshot 2026-11-05.