OpenAI has started offering GPT-6 Rosalind Research, a version of its flagship AI model set up for biology and drug research, to a limited set of organizations it has approved. The company announced it on Friday, Oct. 30. It runs on the same underlying model as GPT-6 Astra, which OpenAI calls its flagship reasoning model: a model that works through a problem step by step before it answers. OpenAI says it did no extra training for this release. OpenAI says it is designed to help with some research topics that GPT-6 Astra may refuse, work with legitimate uses that could also be misused. OpenAI says it is designed to still refuse requests that would help make biological weapons.
Before granting access, OpenAI says it checks whether an applicant has a legitimate life sciences or related mission, a credible use for the model, proper oversight of its research and the ability to control who uses it. Approved customers can use it in ChatGPT's Work section and in Codex, but not in ChatGPT's ordinary chat. Organizations approved separately can also reach it through OpenAI's API, a connection that lets developers plug a service into their own apps, but for now only for their own internal research tools, not for products they offer customers. In large business accounts it starts in a two-week preview that is switched off until an administrator turns it on. OpenAI requires customers to have a qualified person review the model's answers before those answers shape experiments, publications or other important decisions.
On OpenAI's price list for developers, its standard price is $15 per million tokens sent to it, the small chunks of text, often parts of words, that AI models read and write, and $75 per million it writes. Both are one and a half times the $10 and $50 listed for GPT-6 Astra. On 94 research tasks submitted by working life scientists, OpenAI says it averaged 61.3% at its highest reasoning setting, up from 44.0% for the previous version, GPT-5.5 Rosalind Research.
Under its own system for tracking AI abilities that could cause severe harm, OpenAI places the model at "High" for biology and chemistry, below "Critical". On four tests of skills that could help an expert design a dangerous new threat, such as predicting how changes to a protein affect what it does, it reached none of OpenAI's "Critical" thresholds. OpenAI says those thresholds are early estimates that may change, and that the tests cover specific skills rather than everything that could lead to harm. On four tests of hands-on lab know-how, such as fixing a failed experiment, it cleared OpenAI's "High" mark on three once its refusals were counted as correct answers, which is why OpenAI keeps it in that category. OpenAI also treats it, like GPT-6 Astra, as "Critical" for cybersecurity.
OpenAI also ran its own checks, built to be difficult, of how the model and its safeguards answer prompts about biological weapons. On the most severe prompts, the set OpenAI says it mainly weighed for this release, 98.1% of answers met OpenAI's safety bar, against 99.8% for GPT-6 Astra and 88.6% for the previous Rosalind version. On a second set of weapons-related prompts, 79.0% did, against 99.8% for Astra and 18.3% for the previous version. OpenAI says an answer that misses the bar is not necessarily unsafe.
These figures all come from OpenAI's own testing. OpenAI says it tested how well its monitoring catches misuse on the previous Rosalind version and is now repeating that test for the new one. It is keeping the details private because they describe its safeguards, though it says it shared them with its own safety advisory group.