# GPT-Rosalind-6 GPT-Rosalind-6 is [[OpenAI]]'s model for biology, drug discovery and translational medicine, sold as GPT-Rosalind-6 Research to vetted scientists, research institutes, companies and government partners. OpenAI published its system card addendum on 30 October 2026. The name most likely honors Rosalind Franklin. The most interesting line in the document is this one: "GPT-Rosalind-6 uses the same weights as [[GPT-6 Astra]], our flagship reasoning model." And: "We did not add training data or perform additional post-training for this release." So GPT-Rosalind-6 is Astra with a different access policy and different safeguards. ## What's actually different - **Who can use it.** Access is limited to approved customers. OpenAI checks that the applicant has a legitimate life sciences mission, a credible use case and proper research governance. Companies go through business verification and compliance screening; governments get a separate path for biosecurity, biosafety, public health and biodefense work - **Where it refuses.** It's tuned to support more dual-use biology research than the general models, while still refusing requests that would "meaningfully enable severe biological weaponization". On a biorisk knowledge test it refused 9.4% of the time, against 29.2% for Astra - **Cyber.** It inherits Astra's cyber safeguards with "a more conservative cyber refusal boundary". Real-time cyber blocking stays on - **Customer duties.** Enterprises must restrict access to authorized users, apply least privilege, and require qualified human review before an output informs an experiment, a publication or a decision How do you change refusal behavior without post-training? The addendum doesn't say much. It mentions that biological training is kept separate from classifier training, which suggests the difference lives in the classifiers and filters around the model ([[AI Guardrails]]) rather than in the weights. HN commenters guessed the same: a packaged system prompt plus a tuned safety filter. ## Numbers Against the previous version, GPT-Rosalind-5.5, both at the highest reasoning budget: | Benchmark | GPT-Rosalind-6 | GPT-Rosalind-5.5 | |---|---|---| | MedChemBench v7 (99 problems) | 49.6% | 28.9% | | GeneBench Pro v13 (129 problems) | 37.7% | 16.7% | | LifeSciBench Gold (94 problems) | 61.3% | 44.0% | Big jumps, but they come from Astra's weights, not from anything Rosalind-specific. And OpenAI says the tests were "deliberately created to be difficult", built around cases where existing models did badly, GeneBench used a configuration that differs from production, and six LifeSciBench problems were dropped. ## Preparedness ratings Under OpenAI's Preparedness Framework: - **Biological and chemical: High**, "below the threshold for Critical". After adjusting for refusals, it passed the High threshold on biorisk knowledge (85.0% vs an 80% threshold) and on three of four virology troubleshooting tests. None of the four Critical evaluations reached its threshold - **Cybersecurity: Critical**, carried over from Astra - **AI self-improvement: below High**, carried over from Astra Note the oddity: Astra itself scores higher than Rosalind on refusal-adjusted biorisk knowledge (90.0% vs 85.0%). Astra knows the same things; it just refuses more often. That's the whole product in one number. Two safety rows go the wrong way compared to Astra: circumventing a restriction after a warning (21.7% vs 18.9%) and unintended engagement with messages from external agents (2.5% vs 0.0%). And biosecurity red-teaming was still in progress when the card was published. ## Reception The Hacker News thread was mostly skeptical: - If the weights are the same, what are you paying for? Mostly access and fewer refusals - A "card" that runs to hundreds of pages, and yet another confusing OpenAI model name - Who decides who gets access? Some called it undemocratic for a for-profit company to pick which scientists get the stronger tool; others answered that companies have always picked their customers - Is the bottleneck even here? A commenter from a biotech startup said the hard part is manufacturing biologics; another said it's clinical trials ## My take I think this is where frontier AI is heading: one set of weights, several products defined by who you are and what you're allowed to ask. It's the same pattern as Astra's advanced cyber capabilities going first to the US government and "trusted organisations". Capability is no longer the differentiator between tiers; permission is. That's probably the sane way to handle dual-use domains. But it also means the safety story now rests on vetting processes and classifiers we can't inspect, run by a company that decides alone who qualifies. I'd like to see the access criteria and the refusal policy published as clearly as the benchmark tables. ## References - OpenAI, "GPT-Rosalind-6 System Card Addendum" (2026-10-30): https://deploymentsafety.openai.com/gpt-rosalind-6 - Hacker News discussion: https://news.ycombinator.com/item?id=50337503 ## Related - [[GPT-6 Astra]] - [[OpenAI]] - [[AI Safety]] - [[AI Guardrails]] - [[AI Frontier Model]] - [[AlphaFold]]