AttuneBench - Pairwise Accuracy: leaderboard
Metric: Pairwise accuracy (%) at predicting the participant's preferences among three response variants per turn (original, preference-informed alternate, participant-written golden), Default mode (conversation only), 200 conversations; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.7 | 64.6 |
| 2 | Claude Opus 4.6 | 63.7 |
| 3 | MiMo-V2-Pro | 59.8 |
| 4 | Claude Haiku 4.5 | 56 |
| 5 | GPT-5.5 | 53.8 |
| 6 | Mistral Large | 51.7 |
| 7 | Claude Sonnet 4.6 | 51.2 |
| 8 | Gemini 3.1 Pro (Preview) | 50.3 |
| 9 | GPT-5.4 | 47 |
| 10 | Grok 4 | 46.1 |
| 11 | Qwen 2.5 72B | 45 |
Interactive version: theaggregate.ai/benchmark?slug=attunebench-pairwise-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-07.