REFUTE: leaderboard
Epistemic-honesty and scientific-reasoning benchmark (BGPT) scoring models on a composite Truth score that blends accuracy and calibration across four question categories (higher is better).
Metric: Truth Score (accuracy + calibration, higher is better). Source: huggingface.co. Saturation forecast: Around November 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.20 | 74.5 | #176 |
| 2 | Claude Opus 4.7 | 73.4 | #51 |
| 3 | Claude Opus 4.6 | 73.2 | #64 |
| 4 | Grok 3 Mini | 72.2 | #281 |
| 5 | Grok 4.3 | 71.5 | #192 |
| 6 | Gemini 3.1 Pro (Preview) | 71.2 | #59 |
| 7 | Kimi K2.6 | 71 | #109 |
| 8 | GPT-5.2 | 69.2 | #114 |
| 9 | GPT-5.6 Sol | 67.8 | #20 |
| 10 | Gemma 4 31B | 67.2 | #202 |
| 11 | GPT-5.4 | 64.3 | #84 |
| 12 | DeepSeek V4 Pro | 62.9 | #107 |
| 13 | Grok 4.1 Fast | 59.4 | #219 |
| 14 | GPT-OSS-120B | 57.5 | #344 |
| 15 | Llama 3.3 70B | 46.3 | #579 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=refute · How It Works · Data refreshed daily, snapshot 2026-10-24.