OpenAI details how approved researchers can access GPT-Rosalind-6

OpenAI described the screened route to GPT-Rosalind-6 and published a new biorisk evaluation addendum.

OpenAI's October 30 system card addendum describes GPT-Rosalind-6 Research and the controls it applies when serving it to approved scientists, institutions, companies and government partners. The company says the model uses the same weights as GPT-6 Astra. Its deployment is designed to answer more requests related to legitimate life sciences work while retaining refusals around biological weaponization and other restricted uses. [OpenAI system card addendum](https://deploymentsafety.openai.com/gpt-rosalind-6/biorisk-knowledge)

Access is part of the release

OpenAI says it reviews the applicant's proposed work, organizational governance and ability to restrict model access. The release retains the eligibility boundaries already used for GPT-Rosalind-5.5 and described for the Rosalind Academic Plan. The card also says approved enterprise customers must restrict users to authorized purposes and obtain qualified human review before model output informs experiments, publications or consequential decisions. [Trusted-access requirements](https://deploymentsafety.openai.com/gpt-rosalind-6/biorisk-knowledge)

For a research organization, the practical decision is who can use the system and how their work will be reviewed. A research team can document the question posed, check an answer against the available evidence and record which qualified person approved any subsequent use. Those steps remain important even when the model is able to respond to a wider set of legitimate scientific questions.

What the safety tests measure

The card separates beneficial use from tests intended to detect unsafe assistance. OpenAI says its grading relies on several kinds of examples, including held-out synthetic material, red-team cases and production data. It also notes that its benchmarks and grading can change. A score on the current release therefore describes performance under the reported configuration and cannot be treated as a directly comparable trend against every earlier system card. [Evaluation methodology](https://deploymentsafety.openai.com/gpt-rosalind-6/biorisk-knowledge)

The development is significant because the same base weights can be served under different access and safeguard arrangements. Model capability is one part of that decision. Screening, user permissions, monitoring and human review determine how the capability can be used within an institution. OpenAI's addendum sets out those boundaries for this release. It does not offer general ChatGPT access to the research deployment.