GPT-Rosalind-6: Astra's weights, new rules on who gets in

On October 30, 2026, OpenAI released GPT-Rosalind-6, the latest chapter in its Rosalind line for the life sciences. The line launched in April with controlled access and pharmaceutical partners. The naming isn't consistent: product documents call it GPT-6 Rosalind Research, the API changelog calls it GPT-6 Rosalind, and the identifier is gpt-6-rosalind. The changelog announces “Released GPT-6 Rosalind, bringing specialized biological reasoning with scientific tool use to complex life sciences,” and the system card describes it as a frontier reasoning model for biology, drug discovery and translational medicine. The detail that matters sits in that same system card: the model “uses the same weights as GPT-6 Astra, our flagship reasoning model.” There is no additional training data or post-training for this release. Technically, then, it is not a new model: what changes are the entry rules, the refusal boundaries and the governance required of customers.

A note on method that applies to the whole article. Information attributed to OpenAI's Help Center is taken from Forkast's reporting of it. All benchmarks cited are internal to OpenAI and have not been independently verified. Some values in the system card's tables were not extracted, so they don't appear here.

The only way in is the 'trusted access' program: OpenAI reviews applicants' intended uses, security controls and governance. Non-government applicants go through business verification and compliance screening. Government bodies must instead work on approved use cases in biosecurity, biosafety, health preparedness or biodefense. The system card asks customers for three things: restrict use to authorized users, apply least privilege and insider-risk controls, and require qualified human review before results guide experiments, publications or significant decisions. The Discovery and Research tiers are also separate, and the Help Center says so plainly: “Discovery approval does not grant Research access.” The model is available in Work and Codex. The API requires a separate approval, and it is not in the Chat tab. In Enterprise workspaces it starts with a two-week Admin Preview, off by default. As for usefulness, the Help Center places it for now in “early discovery and hypothesis generation.” Two gaps remain. Pricing is unconfirmed: all we know is that rates differ from those of the Rosalind 5.5 models. And there is no list of organizations already approved for the Research tier.

Under the Preparedness Framework, the system card rates biological and chemical capability as 'High', below the Critical threshold. In cybersecurity the rating is 'Critical', carried over from the GPT-6 Astra assessment; in AI self-improvement it stays below 'High'. The biological safeguards are tuned to refuse requests that “meaningfully enable severe biological weaponization”, essentially those that would give concrete help toward developing biological weapons with severe consequences. On cyber, Astra's standard protections remain, including real-time blocking, with a more conservative refusal boundary. OpenAI reserves the right to limit or revoke access in case of misuse.

In internal tests at maximum reasoning budget, the comparison with GPT-Rosalind-5.5 looks like this: 49.6% vs 28.9% on MedChemBench v7, 37.7% vs 16.7% on GeneBench Pro v13, 61.3% vs 44.0% on LifeSciBench Gold. On biological risk knowledge the indicative threshold is 80%. Here cons@32 is 78.3% for GPT-Rosalind-6, 79.2% for 5.5 and 63.3% for GPT-6 Astra, with refusal rates of 9.4%, 1.8% and 29.2% respectively. Adjusted for refusals, the scores rise to 85.0% for GPT-Rosalind-6 and 90.0% for Astra. After this adjustment the model exceeds the indicative 'High' thresholds in three of four biological evaluations, while no 'Critical' capability evaluation reached its threshold; the 'Critical' cybersecurity rating, as noted, is carried over from GPT-6 Astra. The system card does not explicitly explain how weights identical to Astra's produce higher biological scores. Fewer refusals and the scientific tools are the most likely explanation, but that remains a hypothesis.

What strikes me is that this news is more an admissions policy than a model: same weights, different doors. A system rated 'High' in biology arrives through verified access rather than an open release, so its safety depends not only on the model's safeguards but also on who gets past the screening. If OpenAI ever makes public who has been admitted to the Research tier, that will be the page to read most carefully. — Olya