OpenAI has released GPT-Rosalind-6, a frontier reasoning model for biology, drug discovery, and translational medicine that more than doubled its predecessor's score on OpenAI's GeneBench Pro benchmark. The model, published October 30 alongside a system card addendum on OpenAI's Deployment Safety Hub, is available only to vetted organizations through OpenAI's trusted-access program.
GPT-Rosalind-6 runs on the same weights as GPT-6 Astra, OpenAI's flagship reasoning model. OpenAI said it added no training data and did no additional post-training for this release. What changes is the refusal behavior: GPT-Rosalind-6 turns down fewer dual-use biological queries than GPT-6 Astra, while still refusing requests for biological weaponization and keeping refusals on other safety topics.
On MedChemBench v7, a 99-problem test, GPT-Rosalind-6 posted a mean rubric score of 49.6%, up from 28.9% for GPT-Rosalind-5.5. Its single-attempt success rate on GeneBench Pro v13 rose to 37.7% from 16.7% across 129 problems, though OpenAI notes that run used a benchmark-specific configuration that differs from production. On LifeSciBench Gold, the average score climbed to 61.3% from 44.0% across 94 problems. All three results are OpenAI's own; no independent evaluation has been published yet.
Key facts
- 01OpenAI released GPT-Rosalind-6 on October 30, built on GPT-6 Astra's weights with no added training data or post-training.
- 02MedChemBench v7 mean rubric score rose to 49.6% from 28.9% for GPT-Rosalind-5.5.
- 03GeneBench Pro v13 single-attempt success climbed to 37.7% from 16.7%; LifeSciBench Gold rose to 61.3% from 44.0%.
- 04Access is limited to vetted organizations through trusted access; the API model ID is gpt-6-rosalind.
- 05Rated High for biological and chemical capability, below Critical; none of four Critical evaluations met its threshold.
The product is called GPT-Rosalind-6 Research, and its API model identifier is gpt-6-rosalind. OpenAI offers it to qualified scientists, research institutes, companies, and government partners, and reviews each applicant's proposed uses, security controls, and governance before granting access. Eligibility rules are unchanged from GPT-Rosalind-5.5, and the model is not open to general ChatGPT subscribers or general API developers.
Approved enterprises must limit access to authorized users with a legitimate need, apply least-privilege, onboarding, and insider-risk controls, and require qualified human review before model outputs inform experiments, publications, or consequential decisions. Non-government applicants go through business verification and compliance screening. Government applicants have a separate pathway for approved biosecurity, biosafety, public-health preparedness, and biodefense work.
OpenAI has not published a separate price for gpt-6-rosalind. Its pricing page lists the existing gpt-rosalind-research and gpt-rosalind-discovery models at $5 per million input tokens and $25 per million output tokens, half the $10 and $50 rates for standard GPT-6 Astra.
Under OpenAI's preparedness classifications, GPT-Rosalind-6 is rated High for biological and chemical capability, below the Critical threshold. It inherits GPT-6 Astra's Critical rating for cybersecurity and sits below High for AI self-improvement. On TroubleshootingBench it answered 491 of 832 questions correctly, against an indicative High threshold of 36.4%. On a biorisk knowledge test, its refusal-adjusted score was 85.0%, above the 80% threshold, compared with 90.0% for GPT-6 Astra.
None of OpenAI's four Critical-level biology evaluations met its indicative threshold. Those tests cover AAV capsid packaging prediction, SHP2 protein function prediction, a coronavirus-ACE2 cell-entry screen, and phage-plasmid co-evolution. OpenAI describes the Critical thresholds as preliminary estimates that may change.
OpenAI lists four safeguard layers: trusted access, model-level boundaries against harmful biological and cyber assistance, monitoring and enforcement, and contractual security requirements. On cyber, GPT-Rosalind-6 keeps GPT-6 Astra's real-time blocking and adds a more conservative refusal boundary. In a historical jailbreak regression test with real-time blocking disabled, it scored 14.3% against 15.7% for GPT-6 Astra, where lower is better. OpenAI can restrict or revoke access after approval if it detects misuse.
GPT-Rosalind first launched in April 2026 as a research preview for qualified U.S. Enterprise customers, with Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific named among its users. OpenAI said the biosecurity red-teaming it completed for GPT-Rosalind-5.5 is now being repeated for GPT-Rosalind-6, with details withheld.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
