ConfidenceBench - Accuracy: leaderboard
Plain accuracy on the same 200 ConfidenceBench multiple-choice questions, published alongside the calibration scores so that knowing the answer and knowing whether you know it can be read apart. Higher is better. Source: arxiv.org/abs/2607.20526.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 82.7 | #64 |
| 2 | GPT-5 | 76.6 | #105 |
| 3 | Claude Opus 4.6 | 76 | #67 |
| 4 | GPT-5 (High) | 75.8 | #105 (GPT-5) |
| 5 | GPT-5 (Low) | 74.9 | #105 (GPT-5) |
| 6 | Claude Opus 4.5 | 74.5 | #87 |
| 7 | GPT-5 Mini | 72.1 | #193 |
| 8 | Claude Sonnet 4.5 | 70.8 | #148 |
| 9 | GPT-5 Nano | 70.5 | #439 |
| 10 | Gemini 2.5 Pro | 69.4 | #159 |
| 11 | Claude Sonnet 4.6 | 69 | #94 |
| 12 | Claude Opus 4.1 | 67.3 | #146 |
| 13 | Gemini 2.5 Flash | 60.6 | #255 |
| 14 | Claude Sonnet 4 | 57.8 | #207 |
| 15 | Gemini 3.1 Flash Lite | 55.3 | #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-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.