OpenAI Releases GPT-6 Rosalind: Astra Weights, Trusted Access for Life Sciences Research

At a glance

The model landed in the API changelog as gpt-6-rosalind and in a same-day GPT-Rosalind-6 System Card Addendum. Access is limited to organizations approved for Rosalind Research through OpenAI’s trusted-access program—available in Work and Codex, with API use requiring separate approval. It is not available in the Chat tab.

The release date is October 30, 2026, from the system card and the API changelog. The API model ID is gpt-6-rosalind, from the changelog and the Help Center. The weights are the same as GPT-6 Astra, with no added training or post-training for this release, from system card §2. Access is trusted access with Rosalind Research approval, in Work and Codex and in the API where separately approved, and not in Chat, from the changelog and the Help Center. The refusal boundary is fewer refusals on dual-use biology, while still refusing weaponization, from the system card. Preparedness is Biological and Chemical capability High, below Critical; Cybersecurity Critical; and AI Self-Improvement below High, from the system card. API Standard pricing for short context is $15 / $75 per 1M input / output tokens, and $1.50 cached, from the API pricing page.

Same weights, different deployment

OpenAI did not publish a separate GPT-6 Rosalind launch blog on October 30. The first-party record is the changelog entry, the Deployment Safety Hub addendum dated October 30, 2026, plus Help Center and Model Release Notes updates. Naming varies slightly: Help Center and changelog say GPT-6 Rosalind / GPT-6 Rosalind Research; the addendum titles it GPT-Rosalind-6 / GPT-Rosalind-6 Research. The API ID is consistently gpt-6-rosalind.

The addendum’s core technical claim is blunt: GPT-Rosalind-6 uses the same weights as GPT-6 Astra. OpenAI says it did not add training data or run additional post-training for this release. What changes is the deployment—trusted access, a refusal policy tuned for advanced life-sciences work, monitoring and enforcement, and security requirements for approved customers.

OpenAI states that Rosalind has fewer refusals on dual-use biological queries than general access, while still refusing requests that would meaningfully enable biological weaponization. Versus Astra, the addendum also describes a more conservative cyber refusal boundary, while inheriting Astra’s cyber safeguards including real-time blocking. Under the Preparedness Framework, OpenAI treats the launch as High in Biological and Chemical capability (below Critical), Critical in Cybersecurity, and below High in AI Self-Improvement.

Who gets it

GPT-Rosalind is a family with two tiers. Research covers frontier biological work and includes GPT-6 Rosalind Research and GPT-5.5 Rosalind Research. Discovery covers everyday scientific work on Rosalind 5.5 with stricter Discovery safeguards; Discovery approval does not grant Research access or GPT-6. OpenAI says this GPT-6 release does not broaden eligibility beyond prior Rosalind announcements. Intended users in the addendum include qualified scientists, research institutes, companies, and government partners—not institutions alone.

In Work and Codex, GPT-6 Rosalind is available to approved Research users, with a Codex Life Sciences Research plugin for scientific tools. In the API, gpt-6-rosalind is available only where separately approved, for internal research tools—not customer-facing products. The Chat tab does not include it. Enterprise access is a two-week Admin Preview, off by default.

Approved enterprise customers must limit access to authorized users and apply qualified human review before outputs inform experiments, publications, or consequential decisions. The Rosalind Academic Plan for U.S. university labs does not include API access.

Scores and price (OpenAI first-party)

On OpenAI’s own beneficial-capability benchmarks—same problem sets for both models, highest reasoning budget tested—GPT-Rosalind-6 scored 49.6% on MedChemBench v7 versus 28.9% for GPT-Rosalind-5.5; 37.7% versus 16.7% single-attempt success on GeneBench Pro v13; and 61.3% versus 44.0% on LifeSciBench Gold. These are OpenAI evaluations, not third-party leaderboards. On four Critical biological-design proxies, the addendum says none met indicative Critical thresholds.

API Standard pricing for gpt-6-rosalind is $15 / $75 per 1M short-context input / output tokens ($1.50 cached)—above listed rates for gpt-rosalind-research ($5 / $25) and GPT-6 Astra ($10 / $50). Long-context Standard rates are $30 / $112.50.

Bottom line

GPT-6 Rosalind is not a separately trained frontier model. It is Astra’s weights, redeployed for life sciences with a more permissive dual-use biology boundary, offset by trusted access, monitoring, contractual controls, and continued weaponization refusals. For approved Research organizations, it is live in Work and Codex as of October 30, 2026, with API access only where separately approved—and still absent from Chat.