CompanionHarm (Full Codebook): leaderboard
Metric: Macro-F1 (%; 14-way single-label classification of annotated Replika AI companion utterances into 13 socio-emotional harm categories or no harm, each with its preceding multi-turn context; temperature 0; the annotation codebook's category definitions, boundary cases and decision guidance added). Source: arxiv.org. Saturation forecast: Around 2029. 7 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.5 | 45.3 |
| 2 | Claude Opus 4.7 | 43.4 |
| 3 | Gemini 3.1 Pro (Preview) | 42.8 |
| 4 | Qwen 3 32B | 37.3 |
| 5 | Llama 3.1 70B Instruct | 36.7 |
| 6 | Qwen 3 8B | 29.7 |
| 7 | Llama 3.1 8B Instruct | 13.3 |
Interactive version: theaggregate.ai/benchmark?slug=companionharm-full-codebook · How It Works · Data refreshed daily, snapshot 2026-09-29.