OpenAI has released GPT-Rosalind-6, a new version of its life sciences model aimed at biology, drug discovery and translational medicine. According to the system card addendum published on October 30, it runs on the same weights as GPT-6 Astra, OpenAI’s flagship reasoning model, but ships with a looser refusal boundary on dual-use biology questions. The catch: you can only use it if OpenAI approves you through its trusted-access program.
Key takeaways
- GPT-Rosalind-6 uses GPT-6 Astra’s weights with no extra training data or post-training. What changes is how it is deployed and what it will answer.
- OpenAI’s own benchmarks show large jumps over GPT-Rosalind-5.5, including 49.6% vs 28.9% on MedChemBench v7.
- Access is gated: business verification, compliance screening, security controls and mandatory human review of outputs.
- OpenAI rates the model “High” for biological and chemical capability (below “Critical”) and “Critical” for cybersecurity.
- Pricing for the new model and the full list of product surfaces have not been clearly confirmed yet.
What happened
OpenAI first introduced the Rosalind line in April 2026 with GPT-Rosalind for life sciences research, a limited-access model built for scientists rather than the general public. Early users named at that launch included Amgen, Moderna, the Allen Institute and Thermo Fisher Scientific. GPT-Rosalind-6 is the latest step in that line, offered commercially as “GPT-Rosalind-6 Research.”
The interesting technical detail is what OpenAI did not do. The addendum says the model “uses the same weights as GPT-6 Astra” and that no additional training data or post-training was added for this release. In other words, this is not a new model trained from scratch on biology papers. It is OpenAI’s most capable general reasoning model, deployed to a vetted audience with a different set of rules about what it will and won’t discuss.
That difference shows up clearly in refusal behaviour. On OpenAI’s biosecurity refusal tests, GPT-Rosalind-6 refused sensitive material far less often than GPT-6 Astra and stopped over-refusing benign requests, according to Data Studios’ breakdown of the system card. OpenAI says the model is still tuned to refuse requests that “would continue to meaningfully enable severe biological weaponization.”
The benchmark numbers
OpenAI compared GPT-Rosalind-6 against its predecessor, GPT-Rosalind-5.5, on three internal life sciences benchmarks:
- MedChemBench v7 (99 problems): 49.6%, up from 28.9%
- GeneBench Pro v13 (129 problems, single attempt): 37.7%, up from 16.7%
- LifeSciBench Gold (94 problems): 61.3%, up from 44.0%
Those are big relative gains, but read them carefully. All three are OpenAI’s own benchmarks, each with roughly 100 problems, and no independent evaluation has been published yet. OpenAI also notes that the GeneBench Pro run used a configuration that differs from production. And even the headline score means the model still gets about half of the medicinal chemistry problems wrong. This is a research assistant, not an autonomous drug designer.
Why it matters
Biology is one of the hardest areas for AI labs to serve well. The same knowledge that helps a researcher troubleshoot a viral vector experiment could, in the wrong hands, help someone cause harm. General-purpose chatbots handle this by refusing a lot, which frustrates legitimate scientists. GPT-Rosalind-6 is OpenAI’s clearest attempt yet at a different model: keep the capability, but control who gets it.
OpenAI’s own safety numbers show why it is being careful. Under its Preparedness Framework, the company treats the launch as “High capability in the Biological and Chemical domain, below the threshold for Critical.” The model exceeded indicative High thresholds on three of four High-level biology evaluations after adjusting for refusals, while none of four Critical-level evaluations met their thresholds. OpenAI calls those Critical thresholds “preliminary estimates,” which is worth keeping in mind.
Because it shares weights with GPT-6 Astra, the model also inherits Astra’s “Critical” cybersecurity rating. OpenAI says it adds a more conservative cyber refusal boundary on top of Astra’s real-time blocking. The dual-use problem is not unique to biology. We have covered how the same tension plays out in security, where AI agents now find software bugs faster than humans can fix them.
There is also a business story here. As i-SCOOP argues, Rosalind shows OpenAI selling governed access, not just a model. For pharma companies, that means the contract, compliance and audit terms may matter as much as benchmark scores.
What it means for developers and businesses
If you are a developer at a typical startup, you cannot simply swap this model into your app. It is not available to ChatGPT subscribers or to general API developers. Access runs through OpenAI’s trusted-access program, which the addendum describes this way:
- Non-government applicants go through business verification and compliance screening.
- Government applicants have a separate pathway for biosecurity, biosafety, public-health preparedness and biodefense work.
- Enterprise customers must restrict access to authorized users, apply least-privilege and insider-risk controls, and keep qualified people in the loop. In OpenAI’s words, “Customers must apply qualified human review before model outputs inform experiments.”
- OpenAI monitors usage and can restrict or revoke access for misuse.
Eligibility has not been widened beyond the existing Rosalind program. If your organization already has Rosalind access, the practical questions are which surfaces carry the new model and what it costs. Neither is fully settled in public documentation. OpenAI’s existing Rosalind research models are listed at $5 per million input tokens and $25 per million output tokens, while GPT-6 Astra is listed at $10 and $50, but OpenAI has not clearly said which rate the new model uses. Budget for uncertainty until that is confirmed.
For builders in life sciences tooling, the bigger lesson is architectural. The frontier labs are increasingly shipping one set of weights with different policies for different audiences. If you are building on top of these models, expect gated tiers, mandatory audit trails and human-review requirements to become normal parts of the integration, not afterthoughts. For a broader view of how the main models compare on everyday work, see our Claude vs ChatGPT vs Gemini showdown.
What’s still unclear
A few points deserve caution. Some coverage has linked GPT-Rosalind-6 to large pharma deals, but we could not confirm those in OpenAI’s own materials, so we have left them out. The named partners above come from the original April launch, not this release. OpenAI is also repeating its biosecurity red-team exercise for this model and has not published those results. Finally, some alignment metrics in the addendum are slightly worse than GPT-6 Astra’s, such as a 2.5% rate of unintended engagement with external agent messages versus 0.0% for Astra.
What to watch next
- Independent benchmarks. Until outside researchers test the model, the gains are OpenAI’s claims.
- Pricing and surfaces. Whether GPT-Rosalind-6 bills at Rosalind or Astra rates, and whether it reaches tools like Codex for approved users.
- Red-team results. The repeated biosecurity exercise will show whether looser refusals hold up under pressure.
- Competitor responses. Google DeepMind, Anthropic and specialist labs are all chasing scientific AI. Expect similar gated offerings.
The short version: GPT-Rosalind-6 is less about a new model and more about a new deal. OpenAI is betting it can give vetted scientists a more capable, less restrictive tool while keeping the riskiest knowledge behind identity checks, contracts and monitoring. Whether that balance holds will matter well beyond biology.