OpenAI's biology model is its flagship — with fewer refusals

OpenAI published a system-card addendum on October 30 for GPT-Rosalind-6, a model for biology, drug discovery and translational medicine offered only to approved labs, institutes and government partners.

The twist: it shares its weights with GPT-6 Astra, the flagship reasoning model, with no new training or post-training — the difference is the refusal settings and who is allowed in. Rosalind refuses fewer dual-use biology queries, with a 9.4 percent refusal rate on OpenAI's biorisk test versus 29.2 percent for Astra, while still refusing requests aimed at biological weapons.

Access costs one and a half times standard Astra rates and requires a trusted-access review; approved users must apply human review before outputs shape experiments or publications.