OpenAI released GPT-Rosalind-6 as a biology-focused reasoning model. It shares weights with GPT-6 Astra but targets drug discovery and translational medicine. Access is restricted to trusted partners after a security and governance review. The model refuses biological weaponization requests but allows some dual-use queries. OpenAI rates it High capability in Biological and Chemical domains. It scored higher than GPT-Rosalind-5.5 on MedChemBench, GeneBench Pro, and LifeSciBench Gold. Preparedness evaluations show it meets High capability thresholds in three of four biological tests. It is treated as Critical in Cybersecurity and below High in AI Self-Improvement.
What changed
GPT-Rosalind-5.5 was the prior biology-focused model with lower benchmark scores.
What it unlocks
Qualified scientists can now access a higher-capability biology model for research under trusted access.
- 49.6% MedChemBench v7 mean score
- 37.7% GeneBench Pro v13 success rate
- 61.3% LifeSciBench Gold average score
- 78.3% Biorisk knowledge cons@32
What you need to act on it
- qualified scientist, research institute, company, or government partner status
- trusted access approval after use-case and security review
Sources