OpenAI Offers GPT-Rosalind-6 to Approved Life-Sciences Researchers With Astra Weights and Trusted Access

OpenAI published a GPT-Rosalind-6 system-card addendum on October 30. This November 1 issue reports that original date rather than presenting the issue date as a new launch. OpenAI offers GPT-Rosalind-6 Research to qualified life-sciences researchers through a reviewed, trusted-access process. The model uses the same weights as GPT-6 Astra; the company says it added no training data and performed no additional post-training for this release. The distinction lies in the intended research setting, access conditions, behavior boundaries and deployment safeguards, not in a newly trained set of weights.

The stated uses include biology, drug discovery and translational medicine. OpenAI says Rosalind refuses fewer dual-use biological questions than Astra in general access, so that legitimate advanced research can receive useful answers. It also says the model continues to refuse requests that would meaningfully enable biological weaponization and retains refusals in other safety areas. Fewer refusals should not be read as unrestricted biological assistance or ordinary public ChatGPT availability. Eligible scientists, research institutes, companies and selected government partners must pass an application review that examines their purpose, security controls and operational governance before access is granted.

Capability labels apply to specific risk domains. Under OpenAI's Preparedness Framework, the launch is treated as High in Biological and Chemical capability, below its Critical threshold. For Cybersecurity it inherits the assessment of the parent GPT-6 Astra model and is treated as Critical; for AI Self-Improvement it is below High. Shared weights do not make the deployment rules identical in every domain. A system card can discuss model capability, model-level refusals and system-level controls separately, and a reader should not treat any one of these as a complete substitute for the others.

The addendum says its biological assessment uses four High and four Critical evaluations from the Astra system card. The former examine wet-lab knowledge and troubleshooting; the latter include design capabilities relevant to severe biological threats. OpenAI explicitly limits what this evidence means: the evaluations cover selected capabilities, not a complete test of whether a real-world harm can be caused. Its Critical thresholds are preliminary estimates informed by experts and may change. Some benchmark suites and graders have also been updated, so a score from an older card is not automatically comparable with a score in this one. Where a table states a denominator or counts unsuccessful evaluation attempts as incorrect, that condition belongs with any numerical quotation.

Access review is a substantive part of the release. For non-government applicants, OpenAI describes business verification and compliance screening in addition to assessment of the proposed use and safety posture. A separate trusted-access path exists for selected government work on approved biosecurity, biosafety, public-health preparedness and biodefense use cases. Approved enterprises must restrict access to authorized people with a legitimate need, use least privilege and onboarding controls, and manage insider risk. Qualified human review is required before model output informs experiments, publications or consequential decisions. Model access and data-retention terms are approved separately; receiving one approval does not silently settle the other.

OpenAI describes four components of the safeguard approach: verified trusted access, model-level boundaries against harmful biological and cyber help, monitoring with enforcement, and additional contractual obligations. It says Rosalind retains Astra's standard cyber safeguards, including real-time blocking, and can restrict or revoke access after misuse. These are statements about its own deployment design. The published evaluations do not prove that every attempted misuse will be caught or that a safe-sounding answer is suitable for an experiment. The company says it continues to learn from use and may adjust safeguards as deployment evolves.

The practical interpretation is a narrower, governed research release of a powerful existing model. The addendum supplies evidence about selected capability tests and explains who may apply and which controls must surround use. It is not an independent clinical validation, a drug discovery result, or a license to run unsupervised wet-lab work from model output. A prospective organization should check eligibility, map its authorized users and human-review process, and read both this addendum and the parent Astra card before relying on a particular metric. This publication has not reproduced the model's capability or safety evaluations; all such claims remain attributed to OpenAI.