OpenAI releases GPT-6 Rosalind, a biology model for approved researchers only

OpenAI released GPT-6 Rosalind on October 30, a version of its latest AI model built for biology and drug discovery research. Unlike most of its products, it is available only to research organisations that OpenAI has approved.

Same model, different rules

According to OpenAI's system card addendum, as described in coverage of the launch, GPT-6 Rosalind uses the same underlying weights as GPT-6 Astra, the company's general flagship model, with no additional training.

What changes is how it handles requests. Rosalind has a looser refusal boundary for "dual-use" biology: questions that have legitimate research value but could, in principle, help someone cause harm. In tests reported by one outlet, Rosalind fully answered every benign biology question, compared with about 92 percent for Astra.

That is the reason for the access controls. OpenAI rates the model "High" for biological and chemical capability under its safety framework, but below "Critical". It is offered through what OpenAI calls a trusted-access programme, initially for qualified enterprise customers in the United States. For businesses, it starts with a two-week admin preview and is switched off by default.

What it is for

OpenAI lists target research, genomics and sequencing, protein and sequence analysis, medicinal chemistry, literature synthesis, wet-lab troubleshooting and experiment planning among its intended uses.

The Rosalind series, named after the chemist Rosalind Franklin whose work was central to discovering the structure of DNA, began in April. Because this version uses the same weights as Astra, any improvement over the previous Rosalind model reflects the newer base model rather than specialised biology training.

Why it matters

Life sciences are one of the areas where AI companies see the biggest potential, from finding drug candidates to designing experiments faster. They are also where safety concerns are most acute, because the same knowledge that speeds up vaccine research could help someone misuse a pathogen.

OpenAI's approach, a single model with different rules depending on who is using it, is an attempt to serve researchers without making those capabilities available to everyone. Some commentators note that it means AI companies are increasingly selling access and permissions, not just models, and that research institutions, particularly in Europe, will need governance rules for how they use such tools.

OpenAI says biosecurity red-teaming, in which experts try to get the model to do harmful things, was carried out on the previous Rosalind model before launch and is being repeated for this one.