Atlas

Benchmarks

← All benchmarks

VCT-Refusal (CAIS)

Safety · 2025-11-14

VCT-Refusal is the refusal-rate subset of the Virology Capabilities Test, manually labeled for expert-level virology queries that should be refused because of dual-use risk. SAFE uses it as a bioweapons-assistance risk component in the CAIS AI Dashboard.

Top models (higher is better)

ModelScore
Claude Fable 5100.0
Claude Fable 5.198.9
Claude Opus 5.590.4
Sonnet 4.568.4
GLM-5.266.3
GLM-5.366.3
Muse Spark 1.161.5
Muse Wonder60.2
Opus 4.757.8
Grok 4.2057.8
Grok 4.356.1
Opus 4.550.3
Grok 445.0
Grok 4.1 Fast43.9
Grok 4 Fast43.8
Opus 4.841.2
Sonnet 4.633.2
Opus 4.632.6
Claude Opus 529.4
GPT-5.4 Mini20.7
GPT-5.219.3
GPT-6 Astra17.6
GPT-6.1 Sol16.1
GPT-5.516.0
GPT-5.413.9
Muse Spark Open12.9
GPT-5.6 Sol11.3
Gemini 4 Argon10.7
GPT-5.6 Terra10.7
GPT-57.5
GPT-5.6 Luna7.0
Gemini 3.8 Flash6.4
GPT-5 Nano5.4
Gemini 3.7 Flash4.8
Kimi K2.64.3
GPT-5.12.4
GLM-5.11.6
Grok 4.61.1
GPT-5.4 Nano0.5
Gemini 3.1 Pro Preview0.5
Loading Atlas data…