OpenAI Releases GPT-Rosalind-6 for Approved Biology Research Organizations

What happened

The addendum appears in the system card section of OpenAI's deployment safety hub, dated Oct 30; the model is GPT-Rosalind-6, a life-sciences research model covering scientific literature, biological data, medicinal chemistry and experiment planning.

The weight relationship is the key limit of this release: GPT-Rosalind-6 uses the same weights as GPT-6 Astra, with no additional training data and no further model training; the difference lies in how biological research requests are handled and in the safeguards around access, not in new base training.

Access control is equally tight: applicants must pass a trusted-access review covering their research plans, security controls and governance; approved organizations use the model through Work and Codex, with API access requiring separate approval; Discovery approval does not grant Research access, and Enterprise workspaces start with a two-week administrator preview that is off by default.

Key facts

1. Release form and date: a system card addendum in OpenAI's deployment safety hub, Oct 30, 2026, for GPT-Rosalind-6.
2. Purpose: biology, drug discovery and translational medicine research, for approved scientists, research institutes, companies and government partners.
3. Weights: the same weights as GPT-6 Astra, with no additional training data and no further model training.
4. Medicinal chemistry benchmark: 49.6% at the highest reasoning budget tested, versus 28.9% for GPT-Rosalind-5.5.
5. Genomics and quantitative biology benchmark: 37.7% at the highest reasoning budget tested, versus 16.7% for the predecessor.
6. Applied life sciences benchmark: an average score of 61.3%, versus 44.0% for the predecessor, on tasks contributed by working scientists.
7. Capability ratings: High in the Biological and Chemical domain and Critical in Cybersecurity, relying on GPT-6 Astra's assessments; none of its four Critical biological capability evaluations reached the relevant threshold.
8. Refusal policy: fewer refusals on some dual use biological queries, while still refusing requests for biological weaponization.
9. Access review: applicants must pass a trusted-access review covering their research plans, security controls and governance; buying a subscription alone does not provide access.
10. Channels: approved Research organizations use it through Work and Codex, with API access requiring separate approval; Discovery approval does not grant access to the Research model; Enterprise workspaces start with a two-week administrator preview that is off by default.
11. Result limits: the benchmark scores are OpenAI's results on specific tests and do not establish that the model can safely diagnose patients, choose treatments or deliver a successful new medicine; one biological-data benchmark used a configuration different from the production system.