OpenAI published GPT-Rosalind-6 Research on October 30, 2026, as a model deployment for biology, drug discovery and translational medicine. It uses the same weights as GPT-6 Astra, with no additional training data or post-training for this release, while its research-specific refusal behavior and access controls differ. OpenAI says it responds to some dual-use biology questions that general-access models may refuse and still refuses requests that would meaningfully enable biological weaponization, as described in OpenAI’s system card.
That makes GPT-Rosalind-6 a specialized, approval-based research deployment, not a newly trained biology foundation model. Its published scores measure performance on defined benchmarks, not clinical outcomes.
What changes from GPT-6 Astra
GPT-Rosalind-6 shares GPT-6 Astra’s model weights. OpenAI says it added neither training data nor post-training for this release; the differences are the research deployment, its biological refusal boundaries and its access safeguards.
That setup is intended to let approved researchers ask some dual-use life-sciences questions that a general-access model might refuse. OpenAI says requests that would meaningfully enable biological weaponization remain outside the permitted boundary.
What OpenAI’s biology benchmarks measure
OpenAI reports higher scores for GPT-Rosalind-6 than GPT-Rosalind-5.5 on three life-sciences evaluations. The tests cover specific problem sets and scoring methods, so the results are not evidence of clinical effectiveness or performance across every laboratory task.
| Evaluation and metric | GPT-Rosalind-6 | GPT-Rosalind-5.5 | Test scope and caveat |
| MedChemBench v7, mean rubric score | 49.6% | 28.9% | Same 99 problems; up to nine scored responses per problem at each reasoning budget; scores shown at the highest budget tested. |
| GeneBench Pro v13, estimated single-attempt success | 37.7% | 16.7% | 129 problems; seven responses per model per problem at each reasoning budget; the benchmark configuration differs from production; scores shown at the highest budget tested. |
| LifeSciBench Gold, average score | 61.3% | 44.0% | 94 problems retained from an initial 100 after six were excluded for possible task or evaluation-environment issues; scores shown at the highest budget tested. |
The problem counts and configuration matter when applying these results to research work. In particular, GeneBench Pro’s evaluation setup differs from production, and LifeSciBench Gold’s scores cover the 94 retained problems rather than all 100 originally considered.
How to read the biorisk scores
On a biorisk-knowledge evaluation developed with Gryphon Scientific, OpenAI reports three separate measures: raw consensus across 32 responses, the refusal rate, and a refusal-adjusted consensus score. The adjusted measure counts refusals as correct, so it should not be treated as the same score as raw consensus.
| Measure | GPT-Rosalind-6 | GPT-6 Astra |
| Raw cons@32 | 78.3% | 63.3% |
| Refusal rate | 9.4% | 29.2% |
| Refusal-adjusted cons@32 | 85.0% | 90.0% |
GPT-Rosalind-6 refused less often than GPT-6 Astra on this evaluation, but its refusal-adjusted score was lower: 85.0% versus 90.0%. The raw and adjusted results answer different questions; fewer refusals do not mean a higher adjusted score.
OpenAI’s safety classifications
OpenAI classifies GPT-Rosalind-6 as High for Biological and Chemical capability, below Critical, and Critical for Cybersecurity. It rates the model below High for AI Self-Improvement.
OpenAI says none of its four biological evaluations for Critical capability reached the preliminary indicative threshold. The company also says those evaluations measure specific capabilities, not the full ability to cause biological harm. These are OpenAI’s classifications and assessments.
Who can access it through Work and Codex
OpenAI’s October 30 Enterprise and Edu release note describes a limited rollout for approved Enterprise customers in Work and Codex, not the Chat tab. Both the organization and the individual user need Research approval. The notice describes a two-week Admin Preview with the model off by default; administrators must enable access for approved users. OpenAI’s release note gives those rollout terms.
OpenAI reviews an organization’s research mission, proposed use case, governance, security controls and ability to manage access. Non-government applicants also undergo business verification and compliance screening. Approved customers must restrict access to authorized users and have qualified people review outputs before they inform experiments, publications or consequential decisions.
What the U.S. academic plan includes
For eligible U.S. university labs, OpenAI announced up to 30 Business Standard seats at $0 per user per month for up to 12 months. Access to GPT-Rosalind remains subject to trusted-access approval, and the plan does not include API access. These are terms of the Rosalind Academic Plan, not a general GPT-Rosalind-6 price. OpenAI’s announcement sets out the program’s scope.
What an earlier Codex demonstration shows
An English-language OpenAI demonstration published on April 20, 2026, shows GPT-Rosalind in Codex working from an evidence package about three asthma targets. The workflow uses research-search tools and delegated evidence lanes to assemble a target prioritization. It illustrates an evidence-synthesis process rather than a benchmark result.
