Lifestyle
OpenAI Publishes Safety Details for Biology AI Model GPT-Rosalind-6: Rated "High" Capability, Limited to Vetted Researchers and Partners
On October 30, 2026, OpenAI published a system card addendum for GPT-Rosalind-6, an AI model for biology and drug discovery. It can support advanced research better than openly available ChatGPT models, so it is offered only to vetted scientists, institutions, companies and government partners. All details are OpenAI's own statements.
About 6 min read
What happened
OpenAI's Deployment Safety Hub page shows the GPT-Rosalind-6 system card addendum was published on October 30, 2026. OpenAI says GPT-Rosalind-6 is a frontier reasoning model for biology, drug discovery and translational medicine, the field that turns basic research findings into clinical applications. It is offered as GPT-Rosalind-6 Research to qualified scientists, research institutions, companies and government partners.
OpenAI says the model uses the same weights, meaning the same internal parameters, as its flagship reasoning model GPT-6 Astra. This release added no new training data and no additional post-training (further tuning after initial training). OpenAI also says the model refuses fewer "dual-use" biology questions, where the same knowledge can serve legitimate research or be misused. It is provided only through trusted access and still refuses requests to weaponize biology.
Capability evaluations and risk rating
OpenAI reported results on three life-science benchmarks and compared them with the previous generation, GPT-Rosalind-5.5. OpenAI notes the 5.5 scores use the same benchmark versions and question sets as the new model, so they differ from the figures published when 5.5 launched. The table shows results at the highest reasoning budget tested. A reasoning budget can be understood as the amount of compute the model may use to think while answering.
| Benchmark | GPT-Rosalind-6 | GPT-Rosalind-5.5 |
|---|---|---|
| MedChemBench v7 (medicinal chemistry, 99 questions, average score) | 49.6% | 28.9% |
| GeneBench Pro v13 (genomics, 129 questions, single-attempt success rate) | 37.7% | 16.7% |
| LifeSciBench Gold (applied life sciences, 94 questions, average score) | 61.3% | 44.0% |
The Preparedness Framework is OpenAI's approach to tracking and guarding against frontier capabilities that could cause severe harm. Under it, OpenAI treats this release as "High" capability in biology and chemistry, below the "Critical" threshold. OpenAI explains that "High" evaluations test wet-lab knowledge (hands-on laboratory experiment work) and experimental troubleshooting. "Critical" evaluations test biological design capabilities that could help experts develop highly dangerous novel threats.
OpenAI says that after "refusal adjustment", the model exceeded the indicative threshold (OpenAI's reference score line for judging levels) on three of four "High" capability evaluations. Refusal adjustment counts refused questions as answered correctly, to estimate capability that safety training may mask. None of the four "Critical" evaluations reached the indicative threshold, and results versus 5.5 were mixed. OpenAI also states that the critical thresholds are preliminary estimates based on expert judgment and may be revised, and that these evaluations are not a complete test of the ability to cause biological harm.
On cybersecurity, OpenAI says it treats GPT-Rosalind-6 as "Critical", based on the existing evaluation of GPT-6 Astra; on AI self-improvement it is below "High".
Who can use it and what safeguards apply
OpenAI says only approved customers can use the model. Review covers the applicant's mission, use case, governance and ability to control access. Non-government applicants must also pass business verification and compliance screening. Some government bodies have a dedicated channel for biosecurity, biosafety, public health preparedness and biodefense uses. OpenAI says this release did not loosen eligibility, and previously announced scopes such as the Rosalind Academic Plan remain unchanged.
OpenAI says enterprise customers must limit use to authorized personnel with a legitimate need. They must also grant least-privilege access, meaning each person gets only the access they need, and maintain internal risk controls. Outputs must go through qualified human review before being used in experiments, publications or major decisions. OpenAI adds that it can restrict or revoke access for misuse and will continue monitoring and enforcement.
The table below lists some safety evaluations reported by OpenAI. Biorisk knowledge tests biology-risk-related knowledge; cons@32 means the model answers each question 32 times and the majority answer is scored. A "jailbreak" is a special prompt that tries to get a model to bypass its safety limits. This test ran with live blocking turned off, so it reflects only the model's own behavior.
| Item | GPT-Rosalind-6 | GPT-6 Astra |
|---|---|---|
| Biorisk knowledge cons@32 (unadjusted) | 78.3% | 63.3% |
| Biorisk knowledge refusal rate | 9.4% | 29.2% |
| Historical cyber jailbreak task success rate (lower is better) | 14.3% | 15.7% |
| Warning compliance: rate of bypassing limits after a warning (lower is better) | 21.7% | 18.9% |
Note that OpenAI's own numbers show the model is not better than its parent model on every item. In the warning compliance test, GPT-Rosalind-6 bypassed limits at a rate of 21.7%, higher than GPT-6 Astra's 18.9%. OpenAI also says biosecurity red-teaming (simulating attackers to test safeguards) of GPT-Rosalind-6 is ongoing. Details are withheld because they involve safeguard information, but OpenAI says they were shared with its Safety Advisory Group and other governance bodies.
What it means for general readers
OpenAI says GPT-Rosalind-6 is designed to support legitimate advanced biological research better than openly available ChatGPT models. It can support some dual-use questions that the openly available GPT-6 Astra might refuse. However, these capabilities are offered only to vetted customers, not the general public. OpenAI also says it may update safeguards based on real-world use, safety signals and emerging misuse patterns.
Frequently asked questions
Can the general public use GPT-Rosalind-6?
No. OpenAI says GPT-Rosalind-6 is available only to approved customers, including qualified scientists, research institutions, companies and government partners, who must pass review before getting access.
Is GPT-Rosalind-6 a newly trained model?
No. OpenAI says it uses the same weights as GPT-6 Astra, and this release added no new training data and no additional post-training. The differences lie in deployment controls and refusal behavior.
What does the "High" capability rating mean?
OpenAI's "High" evaluations test wet-lab knowledge and experimental troubleshooting. Its "Critical" evaluations test biological design capabilities that could help experts develop highly dangerous novel threats. OpenAI rates this release "High" in biology and chemistry and says none of the four "Critical" evaluations reached the indicative threshold. This is OpenAI's own classification.
What requests will GPT-Rosalind-6 still refuse?
OpenAI says the model still refuses requests to weaponize biology and keeps refusals on other safety topics. On cybersecurity it uses a more conservative refusal boundary than GPT-6 Astra. It refuses requests that clearly aid unauthorized, destructive or harmful cyber activity, such as deploying malware or stealing credentials or data.
Browse the latest news in this topic
Lifestyle
Sam Altman Says RAM Is Better in Data Centers Than on Your Desk, as the Memory Shortage Pushes Up Device Prices
On a podcast, OpenAI CEO Sam Altman said it is "quite terrible" for the world's limited memory supply to sit in personal computers. The Verge notes that billions spent on chips by big buyers are why laptops, phones and other tech are getting more expensive for ordinary consumers.
Lifestyle
OpenAI, Google, Microsoft and Anthropic Say They Will Adopt One Signal for Opting Out of AI Scraping
The Verge reports that four AI companies say they will adopt a shared signal placed in a site's robots.txt file. Site owners and publishers would no longer need to tell each company separately whether their content may be scraped to train AI.
Lifestyle
AWS: ChatGPT Desktop App Now Usable via Amazon Bedrock in GovCloud, Aimed at High-Security Organizations
AWS says on its official blog that the ChatGPT desktop app can now use OpenAI models through Amazon Bedrock in AWS GovCloud (US). It mainly affects organizations whose security architecture requires GovCloud; AWS also cautions that this setup does not mean a workload is authorized, and organizations must still run their own review.
Lifestyle
Starbucks Wants to Use AI to Recognize Drive-Thru Customers and Show Them Personalized Menus, but It's Still Early
According to The Verge, Starbucks' CEO said the company will explore recognizing participating drive-thru customers as they arrive, showing them personalized digital menus and making payment easier. If this launches, drive-thru menus may no longer look the same to everyone, but the timing is unclear.