ConfidenceBench - Calibration Error: leaderboard
Expected calibration error on ConfidenceBench: the average gap between a stated confidence and the accuracy actually achieved at that confidence. Lower is better. Source: arxiv.org/abs/2607.20526.
Metric: Expected Calibration Error (0-1). Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 0.07 | #106 |
| 2 | Claude Opus 4.6 | 0.07 | #68 |
| 3 | GPT-5 (High) | 0.08 | #106 (GPT-5) |
| 4 | GPT-5 (Low) | 0.08 | #106 (GPT-5) |
| 5 | GPT-5 Nano | 0.08 | #438 |
| 6 | Gemini 3.1 Pro (Preview) | 0.09 | #65 |
| 7 | GPT-5 Mini | 0.09 | #192 |
| 8 | Claude Sonnet 4.6 | 0.1 | #96 |
| 9 | Claude Sonnet 4.5 | 0.11 | #148 |
| 10 | Claude Opus 4.1 | 0.12 | #146 |
| 11 | Claude Opus 4.5 | 0.14 | #88 |
| 12 | Claude Sonnet 4 | 0.17 | #208 |
| 13 | Gemini 2.5 Flash | 0.19 | #254 |
| 14 | Gemini 2.5 Pro | 0.2 | #159 |
| 15 | Gemini 3.1 Flash Lite | 0.34 | #204 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=confidencebench-calibration-error · How It Works · Data refreshed daily, snapshot 2026-11-06.