CarryOnBench - Harmfulness (Benign 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 clarifies the benign intent, counted over the whole conversation, 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 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 2029. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Haiku 4.52.1#583
2Claude Sonnet 42.4#410
3GPT-5.42.7#135
4GPT-OSS-120B2.7#704
5GPT-5 Mini2.8#364
6Llama 4 Maverick3#1110
7OLMo 3.1 32B Instruct3.1#2445
8OLMo 3 7B Instruct3.2#6953
9Gemini 3.1 Pro (Preview)3.6#99
10Qwen 3 235B A22B3.6#628
11Qwen 3 32B3.9#972
12Mixtral 8x7B Instruct3.9#4696
13DeepSeek V3.13.9#511
14Qwen 3 Next 80B A3B Instruct3.9#778

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=carryonbench-harmfulness-benign-clarifications · How It Works · Data refreshed daily, snapshot 2026-11-21.