CarryOnBench - Utility (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. Share of the benign information checklist fulfilled in a single turn where the request states its benign purpose up front, in percent; higher is better.

Metric: Ben-Util (%): share of the benign information-need checklist (about 18 atomic yes/no items per query, built from five models' compliant answers) that the responses fulfil, each item judged separately by Gemini-2.5-Flash, 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); higher is better. Source: arxiv.org. Saturation forecast: Around October 2028. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5 Mini72.1#364
2Qwen 3 Next 80B A3B Instruct70.7#778
3DeepSeek V3.167.6#511
4GPT-5.463.2#135
5Qwen 3 32B61.2#972
6Qwen 3 235B A22B56.8#628
7OLMo 3.1 32B Instruct56.8#2445
8Gemini 3.1 Pro (Preview)52.1#99
9Claude Sonnet 448.8#410
10OLMo 3 7B Instruct48.8#6953
11Claude Haiku 4.545.9#583
12Mixtral 8x7B Instruct40#4696
13Llama 4 Maverick35.7#1110
14GPT-OSS-120B25.2#704

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

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