OpenAI launches GPT-Rosalind-6, a biology-focused reasoning model under trusted access

A new reasoning model built specifically for biology and drug discovery is being released under a trusted access framework. The model, called GPT-Rosalind-6, shares its base architecture with the company's flagship GPT-6 Astra system but is fine-tuned to reduce refusals on legitimate dual-use biological queries. Access is restricted to qualified scientists, research institutes, companies, and government partners who pass a review of their proposed uses, security controls, and governance practices.

The release marks a shift in how frontier AI capabilities are distributed. Rather than a public launch, the company is screening applicants before granting access. The model still refuses requests for biological weaponization and retains standard safety refusals on non-biological topics. Under the company's Preparedness Framework, GPT-Rosalind-6 is classified as High capability in the Biological and Chemical domain - below the Critical threshold - while inheriting GPT-6 Astra's Critical rating in Cybersecurity and a below-High rating in AI Self-Improvement.

What's different under the hood

GPT-Rosalind-6 uses the same weights as GPT-6 Astra. The key difference is a deliberate reduction in refusals for dual-use biological queries that have legitimate research applications. "Rosalind has fewer refusals on dual-use biological queries but is only made available via trusted access," the company said. The model is designed to support work in translational medicine and drug discovery while maintaining hard boundaries around weaponization requests.

The company published an addendum to the GPT-6 Astra system card that details GPT-Rosalind-6's biological capability assessment, additional safety evaluations, and deployment safeguards. Results for the previous generation, GPT-Rosalind-5.5, are included for comparison. The addendum is meant to be read alongside the parent model's broader capability and safety documentation.

How the trusted access process works

The application process evaluates three dimensions: proposed research use cases, existing security controls, and governance practices. This is not an open API. The company reviews each applicant before granting access, a model closer to how sensitive laboratory equipment or biological agents are distributed than how most AI systems are deployed.

The company continues to maintain other deployment safeguards beyond the application review. Those safeguards are not detailed in the source release but are referenced as part of the overall safety architecture. The model also retains full refusals on non-biological safety topics, meaning the reduced-refusal behavior is scoped specifically to biological research contexts.

Where this fits in the risk framework

The Preparedness Framework classification places GPT-Rosalind-6 at High for biological and chemical capabilities - a tier that triggers specific oversight and deployment controls but remains below the Critical threshold that would require more severe restrictions. The Cybersecurity rating of Critical is inherited from GPT-6 Astra, reflecting the parent model's capabilities in that domain.

The AI Self-Improvement rating sits below High, indicating that the model does not pose near-term risks related to recursive self-enhancement. These classifications are part of a broader framework the company uses to evaluate and mitigate risks before deployment. For researchers working at the intersection of AI and biology, understanding these risk tiers is increasingly relevant as models become more capable in specialized scientific domains. Courses like AI for Scientists Courses address how these tools are being integrated into research workflows.

Why this matters for science and research professionals

For researchers in drug discovery, translational medicine, and adjacent fields, GPT-Rosalind-6 represents a model that is less likely to block queries on topics that sit in the gray area between legitimate research and dual-use concern. That reduced refusal rate could accelerate literature review, hypothesis generation, and experimental design work that previous models might have declined to assist with. The trade-off is gated access - researchers at institutions without established security and governance practices may face a higher bar to entry. For those who qualify, the model offers frontier reasoning capabilities tuned specifically for biological domains, with safety boundaries that are enforced through access control rather than blanket refusals.