By Roman Aglaganov
OpenAI has released GPT-Rosalind-6, the second generation of its model for biology, drug discovery and medicine, according to a system card published on 30 October. The first GPT-Rosalind launched in April; partners named then included Amgen, Moderna, Thermo Fisher Scientific and the Allen Institute.
Same model, different rules
GPT-Rosalind-6 is not a separately trained model. It “uses the same weights as GPT-6 Astra”, OpenAI’s flagship, and OpenAI says it added no training data and did no extra post-training for this release. What differs is how it behaves and who can use it. It has fewer refusals on “dual-use” biology questions, research that could help both medicine and weapons, while still refusing requests related to biological weapons.
On OpenAI’s own life-sciences tests, it improves clearly on the previous version. On GeneBench Pro, which tests planning and carrying out analyses of biological data, its single-attempt success rate was 37.7%, against 16.7% for GPT-Rosalind-5.5. On LifeSciBench Gold, built from tasks set by working scientists, it averaged 61.3% against 44.0%.
The risk assessment
Under its Preparedness Framework, OpenAI treats the model as “High” capability for biological and chemical risk, the same rating as GPT-6 Astra. After adjusting for refusals, it exceeded the High thresholds on three of four tests of lab knowledge and troubleshooting. None of the four tests of biological design, which probe abilities that could help an expert create a dangerous new threat, reached the “Critical” threshold. OpenAI calls those thresholds preliminary estimates that may change. It treats the model as Critical for cybersecurity, inherited from GPT-6 Astra.
Who gets access
Access is only through OpenAI’s trusted-access programme. Before approving an applicant, OpenAI checks for a legitimate life-sciences mission, a credible use case, good governance and the ability to control access, and runs business verification and compliance screening for non-government applicants. Approved customers must maintain strict identity and access controls, and OpenAI monitors use for signs of misuse. It says it tested its enforcement through biosecurity red-teaming on the previous model and is repeating that now that GPT-Rosalind-6 is deployed. It is not publishing the details, because they would reveal how its safeguards work.
