CarryOnBench - Harmfulness (Oracle Intent): 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 a single turn where the request states its benign purpose up front, 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 a single turn in which the query states the full benign intent, 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 4.51.5#583
2GPT-5.41.6#135
3GPT-OSS-120B1.6#704
4GPT-5 Mini1.6#364
5OLMo 3.1 32B Instruct1.6#2445
6Qwen 3 235B A22B1.9#628
7OLMo 3 7B Instruct1.9#6953
8Claude Sonnet 42.1#410
9Qwen 3 Next 80B A3B Instruct2.3#778
10Gemini 3.1 Pro (Preview)2.4#99
11Llama 4 Maverick2.4#1110
12Mixtral 8x7B Instruct2.6#4696
13Qwen 3 32B2.8#972
14DeepSeek V3.12.9#511

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-oracle-intent · How It Works · Data refreshed daily, snapshot 2026-11-21.