BMShop Products
← All articles
OpenAI opens GPT-6 Rosalind, its biology AI model, to approved labs
AI News

OpenAI opens GPT-6 Rosalind, its biology AI model, to approved labs

OpenAI launched GPT-6 Rosalind Research on October 30, 2026, a model that shares GPT-6 Astra's weights but answers more biology questions, available only to vetted organisations.

Ammar Rahimi · 31 October 2026 · 3 min read

OpenAI launched GPT-6 Rosalind Research on October 30, 2026. The model is built for biology, drug discovery and translational medicine work. The surprising part: it is not actually a new model. It runs on the same weights as GPT-6 Astra, OpenAI's flagship reasoning model, with no additional training data.

What changed is the refusal boundary and who gets in. Rosalind answers more dual-use biology questions, the kind that help vaccine teams and could also help bad actors. That is why access sits behind a trusted-access program rather than a subscription.

1. Same model, different door

In a system card addendum published October 30, OpenAI states plainly: GPT-Rosalind-6 uses the same weights as GPT-6 Astra, with no added training data and no additional post-training.

The difference is policy. Rosalind answers more dual-use biology questions that the standard model refuses. In a biorisk knowledge test, Rosalind refused 9.4 percent of questions against 29.2 percent for Astra. Those numbers come from i-SCOOP's detailed read of the OpenAI system card.

2. Benchmark gains are large over version 5.5

Three life sciences benchmarks, tested at the highest reasoning budget, comparing Rosalind 6 against Rosalind 5.5:

  • MedChemBench v7 (drug chemistry, 99 problems): 49.6 percent vs 28.9 percent
  • GeneBench Pro v13 (biological data analysis, 129 problems): 37.7 percent vs 16.7 percent
  • LifeSciBench Gold (real research tasks, 94 problems): 61.3 percent vs 44.0 percent

OpenAI itself notes the jump comes from the stronger Astra foundation, not from biology-specific training. A Data Studios analysis, cited by AI Weekly, confirms the performance gain comes from the base model upgrade, not biology fine-tuning.

3. Not for everyday users

GPT-6 Rosalind Research does not appear in the ChatGPT chat tab. It is limited to approved organisations, reachable through Work, Codex and the API (model id gpt-6-rosalind). OpenAI's help centre notes that buying a subscription does not automatically grant access.

In Enterprise workspaces the model starts in a two-week Admin Preview, switched off by default, and admins must enable it manually for approved users. Approved customers must apply qualified human review before model outputs inform experiments, publications or consequential decisions.

4. Risk rating and safeguards

Under OpenAI's Preparedness Framework, this launch is rated High capability in the Biological and Chemical domain, below the Critical threshold. The Critical cybersecurity rating carries over from Astra.

The system card lists four Critical capability evaluations that all fell below the threshold, including AAV capsid packaging prediction and SHP2 protein function prediction. Approved customers must maintain identity controls, least-privilege access and offboarding processes for users who no longer need the model.

5. Pricing and launch partners

According to AI Weekly, pricing is $15 per million input tokens and $75 per million output tokens, one and a half times Astra's $10 and $50. In other words, OpenAI charges a premium for the governance gate and the looser refusal policy, not for extra intelligence.

The GPT-Rosalind introduction page lists Amgen, Moderna, the Allen Institute and Thermo Fisher Scientific among collaborators on the series. AI Weekly also names Novo Nordisk in the launch group.

What it means for you

OpenAI now sells access and governance, not just models. For everyday readers, the takeaways:

  1. No direct access. GPT-6 Rosalind is not in regular ChatGPT. If you work in life sciences, only vetted organisations get in.
  2. This pattern is spreading. OpenAI used the same play for GPT-5.5-Cyber and Astra for Law: a more permissive model in a sensitive field, under monitoring. Expect other fields to follow.
  3. Premium pricing for policy, not capability. The same weights sell at different prices depending on permissions. This is becoming a market norm.
  4. For regular users, standard ChatGPT on Astra remains unchanged. There is nothing to switch today.
  5. If your organisation is serious about biotech, OpenAI's trusted-access application is open to eligible organisations worldwide and safety review takes time. Apply early.

Sources