OpenAI Details GPT-Rosalind-6 Safeguards

OpenAI’s GPT-Rosalind-6 is a reasoning model for biology, drug discovery and translational medicine, offered only to approved researchers and organizations through trusted access. OpenAI reviews applicants’ proposed uses, security controls and governance before granting access. The company says the model gives fewer refusals on dual-use biological questions while continuing to refuse biological weaponization requests and other safety topics.

GPT-Rosalind-6 uses the same weights as GPT-6 Astra, OpenAI’s flagship reasoning model; the company says it added no training data and performed no additional post-training for this release. OpenAI classifies Rosalind-6 as High capability in the Biological and Chemical domain, below its Critical threshold, and as Critical in Cybersecurity. It is below High in AI Self-Improvement.

On MedChemBench v7, at the highest reasoning budget tested, GPT-Rosalind-6 scored 49.6% against GPT-Rosalind-5.5’s 28.9% across the same 99 problems. On GeneBench Pro v13, single-attempt success was 37.7% versus 16.7% across 129 problems; that test used a benchmark-specific configuration different from production. On LifeSciBench Gold, the models were compared on the same 94 problems after six were excluded: Rosalind-6 averaged 61.3%, while Rosalind-5.5 scored 44.0%.

OpenAI assessed four evaluations at the High capability level and four at the Critical level. After adjusting High evaluation scores for refusals, Rosalind-6 exceeded the indicative thresholds on three of four tests; its observed scores were higher than Rosalind-5.5 on three tests and lower on Biorisk knowledge. The evaluations measure particular capabilities, not a model’s complete ability to cause biological harm, and OpenAI says the Critical thresholds are preliminary.

None of the four Critical evaluations met its indicative threshold. Rosalind-6 scored 0.529 on AAV capsid packaging prediction against a 0.600 threshold, mean R² of 0.38 on SHP2 protein function prediction against 0.60, and 0.42 on the Coronavirus-ACE2 cell-entry screen against 0.75. Its phage-plasmid co-evolution negative log-likelihood was 13.41, above the threshold of 9.40 or lower. OpenAI also relies on GPT-6 Astra’s helpful-only evaluations and external biological capability assessments.

Approved customers must demonstrate legitimate scientific research with public benefit, governance and safety oversight, and enterprise-grade access controls. OpenAI lists four safeguard components: trusted-access controls, model-level limits on harmful biological and cyber assistance, monitoring and enforcement to detect abuse, and additional contractual security requirements. Rosalind-6 retains GPT-6 Astra’s standard cyber safeguards, including real-time blocking. OpenAI says the release does not broaden eligibility beyond prior announcements, including the Rosalind Academic Plan.

OpenAI’s GPT-Rosalind-6 is a reasoning model for biology, drug discovery and translational medicine, offered only to approved researchers and organizations through trusted access. OpenAI reviews applicants’ proposed uses, security controls and governance before granting access. The company says the model gives fewer refusals on dual-use biological questions while continuing to refuse biological weaponization requests and other safety topics.

GPT-Rosalind-6 uses the same weights as GPT-6 Astra, OpenAI’s flagship reasoning model; the company says it added no training data and performed no additional post-training for this release. OpenAI classifies Rosalind-6 as High capability in the Biological and Chemical domain, below its Critical threshold, and as Critical in Cybersecurity. It is below High in AI Self-Improvement.

On MedChemBench v7, at the highest reasoning budget tested, GPT-Rosalind-6 scored 49.6% against GPT-Rosalind-5.5’s 28.9% across the same 99 problems. On GeneBench Pro v13, single-attempt success was 37.7% versus 16.7% across 129 problems; that test used a benchmark-specific configuration different from production. On LifeSciBench Gold, the models were compared on the same 94 problems after six were excluded: Rosalind-6 averaged 61.3%, while Rosalind-5.5 scored 44.0%.

OpenAI assessed four evaluations at the High capability level and four at the Critical level. After adjusting High evaluation scores for refusals, Rosalind-6 exceeded the indicative thresholds on three of four tests; its observed scores were higher than Rosalind-5.5 on three tests and lower on Biorisk knowledge. The evaluations measure particular capabilities, not a model’s complete ability to cause biological harm, and OpenAI says the Critical thresholds are preliminary.

None of the four Critical evaluations met its indicative threshold. Rosalind-6 scored 0.529 on AAV capsid packaging prediction against a 0.600 threshold, mean R² of 0.38 on SHP2 protein function prediction against 0.60, and 0.42 on the Coronavirus-ACE2 cell-entry screen against 0.75. Its phage-plasmid co-evolution negative log-likelihood was 13.41, above the threshold of 9.40 or lower. OpenAI also relies on GPT-6 Astra’s helpful-only evaluations and external biological capability assessments.

Approved customers must demonstrate legitimate scientific research with public benefit, governance and safety oversight, and enterprise-grade access controls. OpenAI lists four safeguard components: trusted-access controls, model-level limits on harmful biological and cyber assistance, monitoring and enforcement to detect abuse, and additional contractual security requirements. Rosalind-6 retains GPT-6 Astra’s standard cyber safeguards, including real-time blocking. OpenAI says the release does not broaden eligibility beyond prior announcements, including the Rosalind Academic Plan.