CarryOnBench - Harmfulness (No Clarifications): 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 multi-turn conversations where the user only pushes back, rephrases or makes non-contentful replies, 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 multi-turn conversations of 4 to 12 turns with only pushback, rephrasing and non-contentful user turns, 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: Not forecast. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Haiku 4.51.7#587
2Claude Sonnet 41.7#414
3OLMo 3.1 32B Instruct1.8#2509
4GPT-OSS-120B1.9#712
5GPT-5 Mini2#373
6GPT-5.42.1#138
7Llama 4 Maverick2.1#1113
8OLMo 3 7B Instruct2.1#6953
9Qwen 3 235B A22B2.3#633
10Gemini 3.1 Pro (Preview)2.8#102
11Qwen 3 32B3.1#983
12DeepSeek V3.13.1#511
13Mixtral 8x7B Instruct3.3#4692
14Qwen 3 Next 80B A3B Instruct3.3#781

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-no-clarifications · How It Works · Data refreshed daily, snapshot 2026-11-22.