The Pulse
OpenAI Releases GPT-Rosalind-6 for Approved Biology Research
OpenAI released GPT-Rosalind-6 on October 30 for approved life sciences research teams, with access limited to vetted organizations and users. The company rates its biological capability as high, while saying it fell short of its indicative
AI.info Team · Written with AI · fact-checked ·
OpenAI is making GPT-Rosalind-6 more permissive for advanced biological research while classifying its biological capability as high and restricting access to approved organizations. The model launched on October 30 as GPT-Rosalind-6 Research, aimed at scientists working in biology, drug discovery and translational medicine.
The tension is deliberate: OpenAI says the model refuses fewer questions about dual-use biology than its general-purpose counterpart, but still blocks requests that could meaningfully enable biological weaponization. Researchers must pass an organizational review, and OpenAI says customers need governance, security controls and human review of consequential outputs.
GPT-Rosalind-6 clears three high-capability thresholds
OpenAI’s system-card addendum places GPT-Rosalind-6 in the “High” category for biological and chemical capability under its Preparedness Framework, below the framework’s “Critical” threshold. Its tests cover laboratory troubleshooting and biological design tasks, including predictions involving protein variants, virus-host interactions and biological evolution.
After adjusting scores to account for refusals, the model exceeded the company’s indicative High-capability threshold on three of four evaluations. Its observed scores were higher than GPT-Rosalind-5.5’s on three tests and lower on the biorisk-knowledge evaluation. OpenAI says none of the four Critical-capability evaluations met the indicative threshold; results against GPT-Rosalind-5.5 were mixed.
“None of the four Critical capability evaluations met its indicative threshold.”
OpenAI’s GPT-Rosalind-6 System Card Addendum
The company cautions that the tests measure particular capabilities, not the full possibility of biological harm. It also describes the Critical thresholds as preliminary estimates informed by expert judgment, and says they may change as the evaluations improve.
Same GPT-6 Astra weights, no added training
Despite the specialized research offering, GPT-Rosalind-6 is not a separately trained model, according to OpenAI. It uses the same weights as GPT-6 Astra, with no added training data or additional post-training for this release.
The distinction lies in how the model is deployed and what it is permitted to do. OpenAI says Rosalind is tuned to support legitimate advanced biological research, including some dual-use topics that GPT-6 Astra may refuse in general access, while retaining boundaries against assistance that could enable severe biological weaponization.
Access requires a separate approval
OpenAI is rolling GPT-6 Rosalind Research out to a limited set of approved Research customers in Work, Codex and the API. It is not available in Chat. The API model identifier is gpt-6-rosalind, and approval for Rosalind Discovery does not automatically grant access to Research.
In Enterprise workspaces, the model begins in a two-week Admin Preview and is off by default. Administrators can enable it for approved Research users and set access for after the preview. OpenAI’s release notes say customers can continue to use GPT-5.5 Rosalind Research, while Rosalind Discovery remains on version 5.5.
OpenAI keeps the gatekeeping in place
The company says applicants must show a legitimate research mission, a credible use case, appropriate governance and the ability to control access securely. Approved customers must restrict use to authorized people with a legitimate need, apply least-privilege controls and have qualified humans review model outputs before they inform experiments, publications or consequential decisions.
OpenAI also says it will monitor for misuse and can restrict or revoke access. Its addendum says the company is repeating a biosecurity red-team exercise for GPT-Rosalind-6 after deployment; it does not publish the exercise details. The release expands the model available to approved research teams, but leaves its eligibility rules unchanged from earlier announcements.