OpenAI gives approved organizations access to GPT-Rosalind-6, which shares GPT-6 Astra’s neural network and answers more borderline biology questions

OpenAI gives approved organizations access to GPT-Rosalind-6, which shares GPT-6 Astra’s neural network and answers more borderline biology questions

On October 30, OpenAI began giving vetted organizations access to GPT-Rosalind-6 Research, a version for biology. According to its system card, a technical report, it shares the same weights as the flagship GPT-6 Astra. These weights are the numerical values adjusted during training that encode everything the neural network has learned.

In June, OpenAI described the previous version, GPT-Rosalind-5.5, as GPT-5.5 further trained on biology data and trained to answer difficult questions without refusing. For the sixth version, OpenAI writes: “We did not add training data or perform additional post-training for this release.”

Astra’s refusal boundary, the line between questions it answers and those it declines, is adjustable: OpenAI trained it to apply stricter limits for users flagged as potentially high-risk. According to OpenAI, Rosalind-6 has “fewer refusals on dual-use biology questions,” meaning questions whose answers can support both scientific research and misuse. In a test of laboratory knowledge related to biological threats, Astra refused 29,2% of questions and answered 63,3% correctly, while Rosalind-6 refused 9,4% and answered 78,3% correctly.

When refusals are counted as correct answers to assess capabilities concealed by refusals, Astra scores higher: 90,0%, compared with 85,0% for Rosalind-6. Astra also scores higher on three other laboratory knowledge tests after the same adjustment.

OpenAI focuses primarily on a safety evaluation of responses to severe biological weapons requests: 98,1% of Rosalind-6 responses were judged safe, compared with 99,8% for Astra and 88,6% for Rosalind-5.5. For “sensitive” requests, 79,0% of Rosalind-6 responses were judged safe, compared with 99,8% for Astra and 18,3% for Rosalind-5.5.

On OpenAI’s own tests, Rosalind-6 outperformed Rosalind-5.5 at the highest reasoning effort tested: the average medicinal chemistry score rose from 28,9% to 49,6%, and the score on tasks from practicing scientists rose from 44,0% to 61,3%. Because Rosalind-6 shares Astra’s weights, these gains came from Astra itself. This is illustrated by a task that asks models to predict which variants of a viral capsid used for gene delivery will package their cargo: the correlation between predictions and experimental results (where 1 is ideal) was 0,22 for Rosalind-5.5, 0,529 for Rosalind-6, and 0,528 for Astra.

This task is one of four tests for the Critical risk level: they assess whether a model can help an expert design a new, exceptionally dangerous threat. None of the four reached its indicative threshold (0,600 for this test). OpenAI classifies the release as High risk in biology and chemistry, below the Critical threshold: after adjustment for refusals, Rosalind-6 exceeded the thresholds on three of the four laboratory knowledge tests.

Approval, or trusted access, is granted to an organization: OpenAI reviews its scientific purpose and ability to control access. According to OpenAI’s help page, purchasing a subscription does not automatically confer approval. An approved organization must have a qualified specialist review the model’s conclusions before they inform experiments, publications, or consequential decisions.

On September 8, before Rosalind-6 was released, a user on the OpenAI developer forum who identified themselves as a virologist from Taiwan wrote that they were receiving refusals on their personal account: “my university does not have a so-called institutional account to set up trusted access for biology.” The author of an analysis on i-scoop argues that OpenAI is selling permission alongside the model: buyers apply for authorization and accept obligations.