OpenAI is rolling out GPT-6 Rosalind Research to a limited set of approved Research organizations, with availability in Work, Codex and, where separately approved, the API. It is not available in the Chat tab. This is a controlled life-sciences release, not a general ChatGPT launch. OpenAI’s access documentation makes those boundaries explicit.
The accompanying system card addendum, published October 30, 2026, explains the safety rationale: the model supports legitimate biological research, including some dual-use topics, while retaining refusals for biological weaponization.
For research teams, the consequential detail is how approval works. Discovery access, Research access, individual authorization, product availability and activation are separate checks. Clearing one does not automatically clear the others.
Who Can Request Access?
Eligible organizations with Enterprise or Business agreements can request GPT-Rosalind through OpenAI’s trusted-access program. The safety addendum describes the intended Research audience as qualified scientists, research institutes, companies and government partners.
Buying a subscription or joining an existing workspace does not automatically provide access. OpenAI reviews proposed uses, security controls and governance practices before granting it. The published documents describe those review areas without providing a complete scoring rubric or quantitative acceptance thresholds. Sources: access guide, system card addendum.
Academic applicants have a separate route:
- Labs at U.S. universities without an OpenAI Enterprise agreement: The principal investigator or designated lab lead should follow the Rosalind Academic Plan application process.
- Universities with an Enterprise agreement: Applicants should work through their university administrator or OpenAI account team.
- Applicants unsure which agreement applies: OpenAI advises checking before submitting another application.
The Rosalind Academic Plan is a Business-based offer with its own verification requirements. It is separate from ChatGPT Edu and the ChatGPT for Academic Researchers program, and it includes neither API access nor API credits.
An authorized, verified organization member can submit an institutional application without being a workspace owner or administrator. A verified owner or administrator must complete the separate final approval, accept applicable terms, acknowledge pricing and authorize activation for the correct organization. A university email address does not extend one lab’s approval across every department or workspace. Source: application instructions.
I’d establish who owns that final approval before applying. The process explicitly separates permission to submit a request from authority to activate access, which could otherwise leave a successful application waiting on an internal handoff.
Discovery and Research Have Different Approval Boundaries
OpenAI divides GPT-Rosalind access into two tiers with different model availability and biological safeguards.
| Access tier | Intended work | Models and controls |
|---|---|---|
| Discovery | Literature review, biological data analysis, scientific coding, protocol interpretation and experiment planning | Uses the existing Rosalind 5.5 model with Discovery-specific safeguards and stricter monitoring than Research |
| Research | Frontier biological research | Includes GPT-5.5 Rosalind Research and GPT-6 Rosalind Research, subject to the organization’s specific approval and activation |
Discovery approval does not grant Research access. Moving between the tiers requires additional review and approval. Discovery also excludes both GPT-5.5 Rosalind Research and GPT-6 Rosalind Research. Source: tier definitions.
That distinction matters when organizations describe themselves as having “Rosalind access.” The phrase alone says little about which model an employee can use. A Discovery-approved team may have access for everyday scientific work while remaining ineligible to use either Research model.
Even Research approval is not a blanket entitlement to every release or delivery channel. OpenAI says GPT-6 availability remains limited, and the organization’s approval and activation instructions determine the models and products it can use.
Where GPT-6 Rosalind Research Is Available
OpenAI’s product names need careful reading here. Its broader documentation describes GPT-Rosalind access within ChatGPT, but the specific GPT-6 Research instructions distinguish Work from the Chat tab.
| Product or surface | GPT-6 Rosalind Research availability |
|---|---|
| Work | Available to a limited set of approved Research organizations and users |
| Codex | Available to a limited set of approved Research organizations and users |
| OpenAI API | Available where separately approved |
| Chat tab | Not available |
| Amazon Bedrock | GPT-Rosalind is not currently available |
The academic plan’s exclusion of API access also matters: approval through that plan should not be read as permission to build an API integration. Source: availability and plan restrictions.
Enterprise Starts With an Admin Preview
In Enterprise workspaces, GPT-6 Rosalind Research begins with a two-week Admin Preview and is off by default during that period. Administrators can enable it for approved Research users and configure access for after the preview.
Existing Research approval does not bypass this stage.
The practical implication is that approval and availability can arrive at different times. An organization may qualify for Research while its users still lack an enabled GPT-6 model. OpenAI also explicitly warns that passing verification or accepting terms alone does not mean the model is ready to use. Source: activation and Admin Preview rules.
For administrators, I’d treat the preview as a deployment checkpoint: confirm the approved users, the intended workspace and the post-preview settings before telling researchers the model is available.
Why the Biology-Specific Safeguards Matter
OpenAI says GPT-Rosalind-6 uses the same weights as GPT-6 Astra. It added no training data and performed no additional post-training for this release.
That is a consequential detail. The documentation describes a specialized research offering without claiming that OpenAI trained a new biology-specific checkpoint for this version. Access controls and permitted research use are central to the release’s identity. Source: model data and training.
OpenAI says Rosalind has fewer refusals on dual-use biological queries while retaining refusals for biological weaponization and other safety topics. “Dual-use” describes knowledge that can serve legitimate research while also carrying misuse potential. Simply blocking every overlapping topic would constrain useful scientific work; allowing it more broadly makes decisions about users, purposes and access controls more consequential.
Under its Preparedness Framework, OpenAI treats this launch as High capability in the Biological and Chemical domain, below Critical. That is OpenAI’s own risk classification. The addendum says its High-capability biological evaluations examine wet-lab knowledge and troubleshooting.
Read together, the documents describe two complementary controls:
- Model safeguards govern what the system will help with, including retained refusals for weaponization.
- Trusted-access controls govern who can use it and under what organizational conditions.
Frequently Asked Questions
3 questions
1Must the Applicant Be a Workspace Administrator?
No. An authorized, verified organization member can apply, but a verified owner or administrator must complete final approval and authorize activation.
