Atlas

Benchmarks

← All benchmarks

BioTIER Refusal

Safety · 2026-04-14

BioTIER Refusal measures refusal behavior on the BioTIER v1 refusal evaluation: 501 prompts spanning select agents and toxins, biomedical dual-use research of concern, and catastrophe-avoidance biological-risk topics. Scores are refusal rates in percentage points, so higher is safer.

Top models (higher is better)

ModelScore
Claude Fable 5100.0
Claude Opus 5100.0
Claude Fable 5.199.5
Claude Opus 5.598.7
Muse Wonder98.7
Sonnet 4.598.5
Muse Spark 1.197.5
Muse Spark Open96.0
Opus 4.794.7
Opus 4.894.7
Sonnet 4.694.2
Opus 4.693.0
Opus 4.591.7
Gemini 4 Argon87.2
Gemini 3.7 Flash84.4
Gemini 3.8 Flash83.9
GPT-6 Astra79.9
GPT-5.479.7
GPT-578.6
GPT-5 Nano78.6
o375.6
GPT-5.573.4
Gemini 3.1 Pro Preview72.1
GPT-5.6 Sol71.9
Grok 4.2071.6
GPT-5.4 Mini71.4
GPT-5.271.1
O4 Mini71.1
GPT-5.169.6
GPT-5.6 Luna68.6
Grok 4.668.6
Gemini 3 Flash Preview68.1
GPT-5.6 Terra65.6
GPT-5 Mini64.8
GPT-5.4 Nano64.6
Grok 4.1 Fast64.6
o1 (source unspecified)64.6
Grok 4 Fast63.8
Grok 4.363.6
Grok 462.6
Loading Atlas data…