OpenAI Ships GPT-Rosalind-6 for Biology Research Under Trusted Access
OpenAI released GPT-Rosalind-6 Research on October 30, 2026, a biology and drug discovery model that reuses GPT-6 Astra weights and refuses less often on dual-use biological queries, available only to vetted scientists, institutes, companies and government partners.
OpenAI has released a new biology model that is deliberately less guarded than its flagship system, and deliberately harder to get. GPT-Rosalind-6 Research, published on October 30, 2026, is a frontier reasoning model for biology, drug discovery and translational medicine, offered only through a trusted-access program to qualified scientists, research institutes, companies and government partners.
The release is notable for what did not change under the hood. GPT-Rosalind-6 uses the same weights as GPT-6 Astra, OpenAI’s flagship reasoning model. OpenAI says it added no training data and performed no additional post-training for this release. The difference is in deployment: Rosalind has fewer refusals on dual-use biological queries, the class of questions that can serve legitimate research but also carry misuse risk. The company says it reviews each applicant’s proposed uses, security controls and governance practices before granting access, and that other deployment safeguards remain in place.
On OpenAI’s own benchmarks, the new model is a large step up from its predecessor. At the highest reasoning budget tested, GPT-Rosalind-6 scored 49.6 percent on MedChemBench v7, which tests chemical structure interpretation, structure-activity relationships and predictions of drug potency and toxicity, against 28.9 percent for GPT-Rosalind-5.5 on the same problems. On GeneBench Pro v13, which tests planning and carrying out quantitative analyses of biological data, single-attempt success rose to 37.7 percent from 16.7 percent. On LifeSciBench Gold, built from research tasks contributed by practicing life scientists, average scores rose to 61.3 percent from 44.0 percent. OpenAI notes that the Rosalind-5.5 scores use the same benchmark versions and problem sets as the new model, so they differ from figures reported at the earlier launch.
The safety framing is explicit, and it is the part of the release most likely to draw scrutiny. Under OpenAI’s Preparedness Framework, the company treats this launch as High capability in the Biological and Chemical domain, below the threshold for Critical. It says the model is designed to support legitimate biological research, including some dual-use topics, while still refusing requests for biological weaponization and retaining refusals on other safety topics. Approved customers must maintain security controls appropriate for access to a higher-capability life sciences model, and OpenAI says it tested its shared monitoring and enforcement process through a biosecurity red-teaming effort on GPT-Rosalind-5.5 and is repeating that work for GPT-Rosalind-6 now that it is deployed. The company does not publicly disclose those safeguard details, saying they contain specific information about its protections, but says it shared them with its Safety Advisory Group committee and other governance bodies.
The trade-off here is familiar in frontier biology work, but Rosalind-6 states it more plainly than most releases do. The same capability that helps a medicinal chemist interpret a structure-activity relationship or plan a genomics analysis can also help someone work through a dual-use biological question. OpenAI’s answer is not to blunt the model for everyone, but to gate it: fewer refusals, narrower distribution, applicant review, monitoring and customer security requirements. Whether that gate holds depends on details OpenAI has not published, including how it vets applicants, how it monitors use after access is granted, and how it responds when a trusted user crosses a line.
For legitimate researchers, the practical question is access. The system card describes Rosalind-6 Research as available to qualified scientists, research institutes, companies and government partners, but it does not publish pricing, a general availability date or an application timeline in the addendum. Researchers who already work with OpenAI’s life sciences tools will need to go through the trusted-access review, and the company says the specific review process depends on the deployment configuration and the information available to it.
The release also clarifies where biology now sits in OpenAI’s risk picture. A High rating in the Biological and Chemical domain, short of Critical, signals a model the company considers powerful enough to warrant special handling but not so capable that it crosses its own threshold for severe-harm risk. That judgment, like the benchmark gains, comes from OpenAI’s own evaluations. Independent researchers will not be able to reproduce the trusted-access version’s behavior without access, which makes outside verification of both the capability claims and the safeguards difficult by design.
What is clear is the direction of travel. Biology is no longer a side application for general-purpose models. It is getting dedicated releases, dedicated benchmarks and dedicated access regimes, with the most capable versions reserved for vetted users. GPT-Rosalind-6 is the latest example of that split: a flagship model, repackaged with looser refusals on dual-use biology, and wrapped in a gate that is meant to keep it in trusted hands.