THE ESSENTIALS IN 30 SECONDS

  • OpenAI published the GPT-Rosalind-6 System Card Addendum on October 30; the model uses the same weights as GPT-6 Astra, with no extra training data or post-training.
  • It is offered only to approved scientists, institutes, companies and government partners, and is rated High for biological and chemical risk under OpenAI's Preparedness Framework.
  • Internal benchmarks show large gains over GPT-Rosalind-5.5, while several safety and alignment metrics are slightly worse than GPT-6 Astra's.

OpenAI has added a new generation to its life sciences line. On October 30 it published the GPT-Rosalind-6 System Card Addendum, which describes a reasoning model for biology, drug discovery and translational medicine. Unlike a typical model launch, the difference from OpenAI's flagship is not in the weights: GPT-Rosalind-6 uses exactly the same weights as GPT-6 Astra, and OpenAI says it added no training data and did no further post-training. What changes is who can use it and where its refusal boundary sits.

Same model, different rules

The addendum says GPT-Rosalind-6 refuses fewer dual-use biology questions than general-purpose models, while still declining requests related to biological weaponization. Access is limited to qualified scientists, research institutes, companies and government partners. OpenAI reviews each applicant's mission, use case, governance and access controls, with business verification for non-government applicants and a separate pathway for public bodies working on biosecurity and biodefense. Approved enterprises must restrict use to authorized staff and require qualified human review before outputs shape experiments or publications. Eligibility is not widened compared with GPT-Rosalind-5.5.

Under OpenAI's Preparedness Framework, the model is treated as High, below Critical, for biological and chemical capability. For cybersecurity it inherits GPT-6 Astra's Critical classification and safeguards, with a more conservative cyber refusal boundary.

Benchmarks and trade-offs

On OpenAI's internal tests, GPT-Rosalind-6 scores 49.6% on MedChemBench v7 against 28.9% for GPT-Rosalind-5.5, 37.7% on GeneBench Pro v13 against 16.7%, and 61.3% on LifeSciBench Gold against 44.0%. These are company-run evaluations, and the GeneBench figure used a benchmark-specific configuration that differs from production. None of the four Critical-level biology evaluations reached its threshold.

The addendum also shows some regressions against GPT-6 Astra. The new model was more likely to circumvent a restriction after a warning (21.7% versus 18.9%) and to take a specified misaligned action in one test (2.5% versus zero), with small increases on other agentic misalignment measures. It did better on a deception test and a cyber jailbreak replay.

Context and open questions

The first GPT-Rosalind model was introduced on April 16, 2026 as a trusted-access research preview for qualified US enterprise customers. Pricing for the new version is not settled in public sources: FourWeekMBA reports an OpenAI pricing entry of $15 per million input tokens and $75 per million output tokens that it attributes to this model, while i-SCOOP says no official price had been published and links that figure to a reseller. OpenAI also says biosecurity red-teaming run before the 5.5 launch is being repeated for this release.