CarryOnBench - Harmfulness (Benign Clarifications): leaderboard
Metric: Harmfulness (1-5): GPT-5.2 rates each response for harmful or policy-violating content (4 or more is harmful); a conversation scores its worst turn, averaged over conversations, over multi-turn conversations of 4 to 12 turns in which at least one user turn clarifies the benign intent, conversation-level, on CarryOnBench (398 seemingly harmful SORRY-Bench queries with a benign underlying intent; user turns simulated by Gemini-3-Flash); lower is better. Source: arxiv.org. Saturation forecast: Around 2030. 14 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Haiku 4.5 | 2.1 |
| 2 | Claude Sonnet 4 | 2.4 |
| 3 | GPT-5.4 | 2.7 |
| 4 | GPT-OSS-120B | 2.7 |
| 5 | GPT-5 Mini | 2.8 |
| 6 | Llama 4 Maverick | 3 |
| 7 | Olmo 3.1 32B Instruct | 3.1 |
| 8 | OLMo 3 7B Instruct | 3.2 |
| 9 | Gemini 3.1 Pro (Preview) | 3.6 |
| 10 | Qwen 3 235B A22B | 3.6 |
| 11 | Qwen 3 32B | 3.9 |
| 12 | DeepSeek V3.1 | 3.9 |
| 13 | Qwen 3 Next 80B A3B Instruct | 3.9 |
| 14 | Mixtral 8x7B Instruct (v0.1) | 3.9 |
Interactive version: theaggregate.ai/benchmark?slug=carryonbench-harmfulness-benign-clarifications · How It Works · Data refreshed daily, snapshot 2026-10-07.