CarryOnBench - Harmfulness (Turn 1): leaderboard

CarryOnBench (2026) starts from seemingly harmful requests that have a benign purpose and checks whether a model gives the benign information a user needs, and how harmful its answers get, as simulated users clarify or push back. Judge-rated harmfulness from 1 to 5 in the first response to the bare request, taking the worst turn; lower is better.

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 the first response to the bare query, 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: Not forecast. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1OLMo 3.1 32B Instruct1.3#2509
2GPT-5.41.4#138
3GPT-OSS-120B1.4#712
4Claude Haiku 4.51.5#587
5GPT-5 Mini1.5#373
6Claude Sonnet 41.5#414
7Qwen 3 235B A22B1.5#633
8Qwen 3 Next 80B A3B Instruct1.5#781
9OLMo 3 7B Instruct1.5#6953
10Llama 4 Maverick1.8#1113
11Gemini 3.1 Pro (Preview)2#102
12DeepSeek V3.12.2#511
13Qwen 3 32B2.3#983
14Mixtral 8x7B Instruct2.8#4692

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=carryonbench-harmfulness-turn-1 · How It Works · Data refreshed daily, snapshot 2026-11-22.