CompanionBench - English - Calibrated Challenge: leaderboard
CompanionBench capability score for Calibrated Challenge: validate-then-challenge, invitational framing, using the user's own words and anti-sycophancy, averaged over English-condition companionship dialogues by a cross-family LLM judge panel (CompanionBench, arXiv 2608.02046v2, Appendix Table H.1). Mean rubric score on a 1-10 scale; higher is better. Capability means correlate strongly across capabilities.
Metric: Mean judge score (1-10). Source: arxiv.org. Saturation forecast: Around February 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 7.96 | #112 |
| 2 | Claude Opus 4.8 | 7.94 | #39 |
| 3 | GPT-5.5 | 7.93 | #33 |
| 4 | Claude Sonnet 4.6 | 7.82 | #94 |
| 5 | GLM-5.2 | 7.79 | #80 |
| 6 | GPT-5.4 | 7.78 | #83 |
| 7 | Qwen 3.5 397B A17B | 7.72 | #154 |
| 8 | Kimi K2.5 | 7.71 | #152 |
| 9 | GPT-5.1 | 7.7 | #145 |
| 10 | Qwen 3.7 Max | 7.7 | #85 |
| 11 | DeepSeek V4 Flash | 7.69 | #155 |
| 12 | Qwen 3 235B A22B 2507 Instruct | 7.69 | #315 |
| 13 | GLM-5 | 7.68 | #149 |
| 14 | GLM-5.1 | 7.67 | #109 |
| 15 | Kimi K2.6 | 7.59 | #113 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=companionbench-english-calibrated-challenge · How It Works · Data refreshed daily, snapshot 2026-11-05.