Safety benchmarks
2 tracked · updated August 16, 2026
These measure harm avoidance rather than capability: how often recommendations carry potential for severe harm, and how models handle mental-health, self-harm, and emotional-reliance conversations. High scores here are a floor requirement for deployment, not a bragging right, and one of the two boards is a vendor's own internal evaluation, labeled as such.
First, Do NOHARM (v2)
Stanford/Harvard consortium · 1,100 consultation cases- 1
Claude Opus 574.6% - 2
Kimi K374.0% - 3
GPT-5.6 Sol70.1% - 4GPT-5.570.0%
- 5
Gemini 3.1 Pro62.6%
OpenAI Dynamic Mental Health Evaluations
OpenAI · 3 safety metrics- 1GPT-5.5 Instant (June Update)0.991
- 2GPT-5.6 Sol (August)0.981
- 3GPT-5.6 Luna (August)0.977
Which safety benchmarks have current frontier-model results?
2 as of August 16, 2026: First, Do NOHARM (v2) (Claude Opus 5 leads at 74.6%); OpenAI Dynamic Mental Health Evaluations (GPT-5.5 Instant (June Update) leads at 0.991).
The other categories sit on the index: rubric-graded benchmarks, agentic and workflow benchmarks, documentation and coding benchmarks, knowledge and exam benchmarks, composite indices.