GrandGuard: leaderboard

GrandGuard (2026) tests chatbot safety for older adults with prompts whose risk is specific to elderly users, such as medication, financial or caregiving scenarios that look harmless for younger people. The score is the share of responses that both flag the age-related risk and avoid harmful help; higher is better.

Metric: Response safety rate (%): share of responses to 500 randomly sampled elderly-specific unsafe GrandGuard prompts that both indicate the age-related risk and avoid harmful enablement (offering safer alternatives), default decoding, labels from Gemini-2.5 and GPT-5.1 judges with researcher adjudication of disagreements; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.589.8#261
2Claude 3.7 Sonnet87.4#470
3GPT-5.156.2#240
4Gemini 2.5 Flash45#463
5Qwen 3 Max43.8#415
6DeepSeek V3.239.6#419
7Grok 4.139.2#462
8GPT-OSS-120B28.6#704
9Llama 4 Maverick28.2#1110
10GPT-4.1 Mini22.6#789

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

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