OpenAI Releases GPT-Rosalind-6 Biology Model Under Trusted Access
3 min readOpenAI has published the system card for GPT-Rosalind-6, a frontier reasoning model built for biology, drug discovery and translational medicine. The GPT-Rosalind-6 release is notable less for new training than for what it unlocks: the model answers more dual-use biology questions than OpenAI’s consumer models, and the company is only handing it to approved customers.
Same weights, different boundaries
According to the addendum dated October 30, 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 ran no extra post-training. What changed is the refusal policy. In OpenAI’s words, the model “has fewer refusals on dual-use biological queries but is only made available via trusted access.”
That access runs through a tier called GPT-Rosalind-6 Research, offered to qualified scientists, research institutes, companies and government partners. Eligibility matches the earlier GPT-Rosalind-5.5 rules. Companies go through business verification and compliance screening, while governments use a separate pathway for biosecurity, biosafety, public-health preparedness and biodefense work. Enterprise customers must restrict the model to authorized users and apply human review before outputs feed into experiments or publications.
What the benchmarks show
OpenAI reports sizeable gains over GPT-Rosalind-5.5 at the highest reasoning budget. On MedChemBench v7 the new model scored 49.6 percent against 28.9 percent. On GeneBench Pro v13 it solved 37.7 percent of problems in a single attempt, up from 16.7 percent. On LifeSciBench Gold it averaged 61.3 percent versus 44.0 percent.
The safety numbers are more mixed. On a biorisk knowledge test the model scored 78.3 percent with a 9.4 percent refusal rate, compared with a 29.2 percent refusal rate for GPT-6 Astra. OpenAI classifies the model as High capability in biological and chemical domains, below the Critical threshold, and says none of its four Critical-tier biology evaluations met their indicative thresholds. Its cybersecurity rating is Critical, inherited from the Astra assessment, which is why real-time blocking of cyber misuse stays switched on.
Why it matters
Drug discovery is one of the clearest commercial cases for frontier AI, and the same knowledge that helps design a therapy can help design a pathogen. OpenAI’s answer is to split the product: one model, two refusal policies, with the permissive version gated behind contracts, verification and monitoring. The company says it can restrict or revoke access for misuse and is repeating its biosecurity red-teaming now that the model is deployed.
Watch for whether competitors adopt the same trusted-access pattern, and whether regulators treat a contract-gated model with relaxed refusals as adequately controlled. For labs, the practical question is simpler: whether a 20-point jump on medicinal chemistry benchmarks turns into faster results at the bench.