Meta's benchmark table puts Gemini 3.1 Pro at 59.2% on biohazard refusals

Meta's Muse Spark Safety & Preparedness Report, dated May 26th, 2026, lists Gemini 3.1 Pro at a 59.2% refusal rate on SecureBio's BioTIER biosafety benchmark in its summary scorecard, with a second table in the same report showing Gemini 3.1 Pro at 61.5%. Meta's own Muse Spark scored 98.0% refusals, Anthropic's Claude Opus 4.6 scored 95.4% and OpenAI's GPT-5.4 scored 75.4% in the six-model comparison, which the report notes is limited by dual Gemini and GPT figures and unclear competitor configurations.

Meta's benchmark table puts Gemini 3.1 Pro at 59.2% on biohazard refusals Meta's May 26th, 2026 Muse Spark report lists Gemini 3.1 Pro at 59.2% in one table and 61.5% in another on SecureBio BioTIER v1 refusals. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published Primary source: Meta Muse Spark Safety & Preparedness Report https://ai.meta.com/static-resource/muse-spark-safety-and-preparedness-report/ Why it matters BioTIER refusal rates show how often a tested system declines prompts that biosafety experts say should be refused. Meta's six-model table gives operators and policymakers a rare side-by-side view, while its dual Gemini and GPT figures and unclear competitor configurations limit how far the ranking can be generalized. Meta's own safety paperwork contains the sharpest cross-lab comparison in the file. In the Muse Spark Safety & Preparedness Report https://ai.meta.com/static-resource/muse-spark-safety-and-preparedness-report/?ref=runtimewire , dated May 26th, 2026, Meta put Muse Spark at a 98.0% refusal rate on SecureBio's BioTIER benchmark, Claude Opus 4.6 https://runtimewire.com/models/anthropic/claude-opus-4.6 at 95.4%, GPT-5.4 https://runtimewire.com/models/openai/gpt-5.4 at 75.4%, and Gemini 3.1 Pro at 59.2% in its summary scorecard.… When you purchase through links in our articles, we may earn a small commission https://runtimewire.com/affiliate-monetization-standards . This doesn’t affect our editorial independence.