KIDBench (No Cue): leaderboard

KIDBench No Cue scores how child-appropriate LLM answers are to realistic questions from children aged 7 to 11 across ten safety-relevant categories when the prompts give no sign that the user is a child (KIDBench, 2026). A DeepSeek-V4-Pro judge rates content safety, developmental appropriateness, socio-emotional support, moral guidance and boundary-setting from 1 to 5; the board shows their mean, higher is better.

Metric: Total score (1-5): mean of the five child-safety dimensions (content safety, developmental appropriateness, socio-emotional support, moral guidance, boundary-setting), each rated 1-5 by a DeepSeek-V4-Pro judge with the KIDBench rubric and category-specific rules, temperature 0 for responder and judge, on the KIDBench child-safety prompts for children aged 7 to 11 (50 per category, 10 categories), single-turn, no-cue prompts (no sign that the user is a child); higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1DeepSeek V4 Flash3.31#267
2Llama 3.3 70B Instruct3.14#1385
3Qwen 3.6 27B3.14#351
4GPT-5 Mini3.06#364
5Claude Haiku 4.53.03#583
6Qwen 3.5 4B2.94#1241
7Qwen 3 8B2.93#1873
8Gemini 3.1 Flash Lite2.92#371
9Gemma 3 12B2.88#2059
10Gemma 4 31B (IT)2.82#350
11Llama 3.1 8B Instruct2.81#4600
12KIDBench Llama-3.2-3B (checkpoint unspecified)2.79
13Gemma 3 4B2.74#5339

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

Interactive version: theaggregate.ai/benchmark?slug=kidbench-no-cue · How It Works · Data refreshed daily, snapshot 2026-11-21.