Grok 4.6 just beat GPT-5.6 Sol, Claude Opus 5 and every other major frontier model tested on an independent biosecurity benchmark And the reason it won is even more impressive than the #1# ranking LatchBio tested whether models can recognize concealed biological risks without becoming useless for legitimate scientific research Grok 4.6 came out as the most performant model tested when balancing: • Refusing disguised biosecurity attacks • Still helping with legitimate dual-use research • Maintaining strong biological intelligence That balance is insanely difficult A weaker safety approach just blocks anything that looks sensitive Grok actually understands the intent behind the request And LatchBio found that much of this behavior comes from the intelligence and reasoning of Grok 4.6 itself rather than heavily relying on external classifiers or system-level blocks At the same time, Grok remains near the absolute frontier across therapeutics, variant discovery, epigenomics, spatial transcriptomics, single-cell transcriptomics and pathogen surveillance