GPT-Rosalind-6 is OpenAI’s new biology and drug-discovery model, and you can’t just sign up for it. OpenAI published its system card on October 30, 2026, and access is limited to approved scientists, labs, companies and government partners. If you were hoping to try it in ChatGPT, that won’t happen.
Here’s the short version of what matters, then the details.
- Released: system card addendum published October 30, 2026.
- Who can use it: approved customers only, through a trusted-access program.
- Under the hood: the same weights as GPT-6 Astra, with no new training data or post-training, per OpenAI.
- Risk rating: High for biological and chemical capability, below the Critical line.
The headline isn’t a new brain. It’s a new set of rules around an existing one.
What GPT-Rosalind-6 actually is
OpenAI’s addendum says it “did not add training data or perform additional post-training for this release.” So GPT-Rosalind-6 is GPT-6 Astra with different deployment settings. The biggest change is that it refuses fewer dual-use biology questions, the kind a vaccine researcher and a bad actor could both ask.
It still refuses requests that would meaningfully help someone build a biological weapon. OpenAI also says it kept a more conservative boundary for cybersecurity requests and real-time cyber blocking.
This is the second Rosalind release. The first arrived in April 2026, which tells you OpenAI sees life sciences as its own product line, not a side feature. For how OpenAI’s other plans are changing, see our post on ChatGPT Pro usage limits doubling.
The benchmark jump, in numbers
OpenAI compared the new model with GPT-Rosalind-5.5 on its own internal tests. These are company-run results, so treat them as a starting point, not a verdict.
| Test | GPT-Rosalind-5.5 | GPT-Rosalind-6 |
|---|---|---|
| MedChemBench v7 | 28.9% | 49.6% |
| GeneBench Pro v13 (single attempt) | 16.7% | 37.7% |
| LifeSciBench Gold | 44.0% | 61.3% |
Medicinal chemistry and genetics tasks improved the most. Even so, AI Weekly notes the model still fails most of the harder biology benchmarks, so it’s a strong assistant, not an autonomous scientist.
Why the same-weights detail matters
If the model underneath is identical, then the safety work happens in the wrapper: which questions get answered, who is asking, and who is watching. That’s a different bet from training a separate, more careful model.
It also raises a fair question. Anyone with access to GPT-6 Astra’s weights would, in theory, have the raw ability. OpenAI’s answer is that the tuned deployment boundaries and the account checks are the control. Whether that holds up is exactly what outside red-teamers will test.
For working scientists, the upside is practical. Fewer needless refusals on legitimate dual-use topics, like pathogen biology or toxicology, means less time rewording questions and more time on the actual research.
Who gets access, and what they must do
Applicants go through a review of their use case, security controls and governance. Non-government applicants also face business verification and compliance screening. Government agencies use a separate pathway for biosecurity and public-health work.
Approval comes with strings attached. Here’s what OpenAI says approved customers must do:
- Limit use to authorized people and apply least-privilege controls.
- Have qualified humans review outputs before they inform experiments, publications or decisions.
- Accept that OpenAI can restrict or revoke access if it spots misuse.
That human-review rule matters. It puts responsibility on the lab, not the model.
How OpenAI rates the risk
Under OpenAI’s Preparedness Framework, GPT-Rosalind-6 is rated High for biological and chemical capability. None of the four evaluations for the Critical level hit their indicative thresholds. Cybersecurity keeps the Critical treatment inherited from GPT-6 Astra.
One test shows why OpenAI is cautious. On biorisk knowledge, the model scores 78.3%, or 85.0% once refusals are counted, which is above the 80% indicative High threshold, per AI Weekly’s reading of the addendum. In plain terms, it knows enough that OpenAI wants a gatekeeper.
If you follow how labs compare models, our explainer on the Intelligence Index v5 shows why vendor-run scores deserve a second look.
What it costs, and what we don’t know
OpenAI’s addendum doesn’t state a price. One analysis from FourWeekMBA says OpenAI’s pricing page lists a gpt-6-rosalind row at $15 input and $75 output per million tokens, versus $10 and $50 for GPT-6 Astra. That’s the writer’s own match of the row to this model, so confirm it with OpenAI before budgeting.
Still unclear is how many organizations have access, and whether government users pay different rates. OpenAI says biosecurity red-teaming is being repeated for this release, so expect more detail.
Anthropic is working the same space, as our post on Claude AI for science partners shows.
Frequently Asked Questions
Can I use GPT-Rosalind-6 in ChatGPT?
No. It’s offered only to approved customers through a trusted-access program, not to the general public.
Is GPT-Rosalind-6 a new model?
Not exactly. OpenAI says it uses the same weights as GPT-6 Astra, with no extra training. The differences are in refusal settings and access rules.
Why does OpenAI limit who can use it?
It answers more dual-use biology questions than general models do. OpenAI rates it High for biological and chemical capability, so it wants vetted users and human review.
How much does GPT-Rosalind-6 cost?
OpenAI hasn’t published a clear price in its system card. One third-party analysis reads the pricing page as 1.5x GPT-6 Astra, but that’s unconfirmed.
Our take
GPT-Rosalind-6 is less a technical leap than a business and safety experiment: sell the same model at a higher tier, behind identity checks. If it works, expect more gated “pro” versions for sensitive fields. If you’re not a vetted researcher, there’s nothing to do but watch, and the benchmark numbers deserve independent testing before anyone calls this a breakthrough.
Leave a Reply