OpenAI releases GPT-6 Rosalind for biology and drug discovery, with GPT-6 Astra weights and trusted access only
OpenAI has released GPT-6 Rosalind, the newest model in its Rosalind life sciences series, and has published a system card addendum that describes it as "a frontier reasoning model for biology, drug discovery, and translational medicine." The release appeared on October 30, 2026 in the OpenAI API changelog under the model identifier gpt-6-rosalind, with the company stating that it brings "specialized biological reasoning with scientific tool use to complex life sciences research." On the same day the OpenAI Deployment Safety Hub published the GPT-Rosalind-6 System Card Addendum, which names the commercial offering GPT-Rosalind-6 Research and sets out its benchmarks, preparedness classifications and access rules.
The most consequential technical detail is that GPT-6 Rosalind is not a separately trained model. According to the addendum, it uses the same weights as GPT-6 Astra, OpenAI's flagship reasoning model, and no additional training data or post-training was applied for this release. What distinguishes it from the general-access model is the deployment around those weights: a different refusal boundary on dual-use biology, separate safeguards, and an access program restricted to vetted organizations. In OpenAI's words, "Rosalind has fewer refusals on dual-use biological queries but is only made available via trusted access," while it still refuses requests that would meaningfully enable biological weaponization.
Access is limited. The API changelog says the model is available to organizations approved for Rosalind Research through OpenAI's trusted-access program, for internal research by authorized users. The system card lists qualified scientists, research institutes, companies and government partners as the intended users, and it states explicitly that this release does not broaden eligibility beyond what OpenAI had already announced for the Rosalind program.
··········
WHAT GPT-6 ROSALIND IS AND HOW IT RELATES TO GPT-6 ASTRA.
The Rosalind line began in April 2026, when OpenAI introduced GPT-Rosalind as a purpose-built reasoning model for biology, drug discovery and translational medicine and as the first release in a life sciences model series. That first version launched as a research preview in ChatGPT, Codex and the API for qualified customers, with initial access limited to qualified U.S. Enterprise customers and named users including Amgen, Moderna, the Allen Institute and Thermo Fisher Scientific. GPT-6 Rosalind is the successor to GPT-Rosalind-5.5, which serves as the comparison baseline throughout the new system card.
The decision to reuse GPT-6 Astra's weights changes how the release should be read. Any gain over GPT-Rosalind-5.5 reflects the jump from the GPT-5.5 generation to the GPT-6 Astra base model, combined with a deployment configuration tuned for research use, rather than a dedicated round of biology training. The system card makes this visible in its own tables: on several biology evaluations GPT-Rosalind-6 and GPT-6 Astra score almost identically, while the differences between them show up mainly in refusal behaviour and safety metrics.
OpenAI's Rosalind product page describes three access options for the series: Rosalind Workbench, an orchestrated workspace that connects scientific tools and is currently in research preview, as a standalone download; the same Workbench inside Codex for existing Codex users; and the GPT-Rosalind API for integrating life sciences reasoning into research applications and governed workflows, subject to eligibility review and organizational approval. The page lists the research workflows the series is meant for, including sequencing data analysis, evidence synthesis, comparison of biological targets, exploration of proteins and variants, comparison of molecular candidates and experiment planning. The GPT-Rosalind-6 addendum itself does not specify which product surfaces carry the new model, so the API listing in the changelog is the only surface OpenAI has documented for it so far.
··········
BIOLOGY AND DRUG DISCOVERY BENCHMARKS AGAINST GPT-ROSALIND-5.5.
The system card reports three capability benchmarks at the highest reasoning budget OpenAI tested, each comparing the new model with GPT-Rosalind-5.5. The gains are large in relative terms, although the absolute scores show that these remain hard, open-ended tasks on which the model still fails more often than it succeeds in two of the three cases.
........
Benchmark | Problems | GPT-Rosalind-6 | GPT-Rosalind-5.5 |
MedChemBench v7 | 99 | 49.6% | 28.9% |
GeneBench Pro v13 (single attempt) | 129 | 37.7% | 16.7% |
LifeSciBench Gold | 94 | 61.3% | 44.0% |
........
On MedChemBench, a medicinal chemistry benchmark, the score rises by 20.7 percentage points. On GeneBench Pro, the single-attempt success rate more than doubles, from 16.7% to 37.7%, although OpenAI notes that this evaluation used a benchmark-specific configuration that differs from production, which means the figure is not a direct measure of what an approved customer will see. LifeSciBench Gold rises from 44.0% to 61.3%. The problem counts are small, between 94 and 129 items per benchmark, so differences of a few points on any single test should be treated with caution.
The system card also reports ProtocolQA Open-Ended, a test of troubleshooting laboratory protocols, where GPT-Rosalind-6 reaches 42.0% pass@1 against 36.4% for GPT-Rosalind-5.5 and 41.4% for GPT-6 Astra. After adjusting for refusals, GPT-6 Astra scores 45.4%, slightly above the Rosalind model, which suggests that on this task the advantage of the Rosalind configuration lies in answering more often rather than in answering better.
··········
PREPAREDNESS CLASSIFICATIONS AND DUAL-USE BIOLOGY RESULTS.
Under OpenAI's Preparedness Framework, GPT-Rosalind-6 is classified as High in Biological and Chemical capability, below the Critical threshold; Critical in Cybersecurity; and below High in AI Self-Improvement. The cybersecurity rating follows from the shared GPT-6 Astra weights rather than from anything specific to the life sciences configuration.
The biological results show a model that clears several of OpenAI's indicative High thresholds. On a biorisk knowledge evaluation scored with cons@32, it reaches 78.3% with a 9.4% refusal rate and 85.0% after refusal adjustment, against an indicative High threshold of 80%. On TroubleshootingBench it answered 491 of 832 items correctly, roughly 59%, against an indicative High threshold of 36.4%. On a multi-select virology troubleshooting evaluation it exceeds the 31% indicative threshold after refusal adjustment. OpenAI summarizes the comparison with GPT-Rosalind-5.5 as mixed.
The four Critical-level biology evaluations are the clearest evidence for OpenAI's position that the model stays below the top tier of risk. None of them met its indicative threshold.
........
Critical evaluation | Metric | GPT-Rosalind-6 | GPT-6 Astra | Threshold |
AAV capsid packaging prediction | Spearman correlation | 0.529 | 0.528 | 0.600 |
SHP2 protein function prediction | Mean R² | 0.38 | 0.40 | 0.60 |
Coronavirus-ACE2 cell-entry screen | Composite score | 0.42 | 0.45 | 0.75 |
Phage-plasmid co-evolution | Neg. log-likelihood (lower is better) | 13.41 | 12.58 | ≤ 9.40 |
........
The near-identical AAV and SHP2 scores for GPT-Rosalind-6 and GPT-6 Astra are consistent with the shared weights. On the phage-plasmid evaluation, GPT-Rosalind-5.5 actually performed better, with a negative log-likelihood of 11.81, than either GPT-6-based configuration.
··········
SAFEGUARDS AND HOW THE REFUSAL BOUNDARY DIFFERS FROM GPT-6 ASTRA.
The practical difference between the two models built on the same weights is captured in OpenAI's biosecurity refusal benchmarks. GPT-Rosalind-6 scores lower than GPT-6 Astra on the categories designed to measure refusals of sensitive material, which is the intended effect of a research configuration, and higher on benign biology requests, where general-access models tend to over-refuse.
........
Biosecurity benchmark | GPT-Rosalind-6 | GPT-6 Astra |
Biosecurity - Sensitive | 79.0% | 99.8% |
Select Sequences | 83.8% | 95.6% |
Dual-Use | 84.6% | 97.8% |
Biosecurity - Benign | 100% | 91.6% |
Chemical Misuse - Severe | 100% | 100% |
........
OpenAI pairs this looser boundary with several layers of control. The model is tuned to refuse or respond safely to requests that would meaningfully enable severe biological weaponization, and the company says that this model-level training is kept separate from the training of its classifiers and offline mitigations, so that the safeguards do not share the same failure modes. On cybersecurity, GPT-Rosalind-6 inherits GPT-6 Astra's safeguards, including real-time cyber blocking, and applies a more conservative refusal boundary than the flagship. In a cyber jailbreak regression test, it recorded 14.3% task success against 15.7% for GPT-6 Astra, where lower is better.
After access is granted, OpenAI says it monitors for misuse and can revoke access. A biosecurity red-team exercise conducted on GPT-Rosalind-5.5 is being repeated for GPT-Rosalind-6, with details withheld. The alignment results are broadly in line with GPT-6 Astra: Blocker Deception is 9.5% against 12.9% for Astra, the OAI-repo Sabotage v2 evaluation shows 0% in both unguided and guided settings, and Unintended Engagement with External Agent Messages rises to 2.5% from 0.0% for Astra. OpenAI notes that these comparisons are descriptive, since versions, settings and grading vary across runs.
··········
WHO CAN ACCESS GPT-6 ROSALIND AND WHAT IT COSTS.
GPT-6 Rosalind is not available to ChatGPT subscribers or to API developers in general, regardless of plan. Approval runs through OpenAI's trusted-access program, which reviews an applicant's life sciences mission, use case, governance and ability to control access. Non-government applicants go through business verification and compliance screening; government applicants have a separate trusted-access pathway for approved biosecurity, biosafety, public-health preparedness and biodefense work.
Approval also carries operating obligations. The system card lists least-privilege access, onboarding, insider-risk controls, identity and access management, and qualified human review of model outputs before they inform experiments, publications or decisions. For research organizations, the last requirement matters most in practice: the model is positioned as an assistant whose outputs must be checked by qualified scientists, not as an autonomous decision-maker in the lab.
On pricing, the API changelog links GPT-6 Rosalind to the life sciences section of OpenAI's pricing page. As listed on the official pricing page on October 31, 2026, that section shows two Rosalind models, gpt-rosalind-research and gpt-rosalind-discovery, but no separate row for gpt-6-rosalind. The page also notes that eligible organizations continue to receive the latest GPT-Rosalind models as they are released, and that cache-write pricing does not apply.
........
Model (official pricing page) | Input per 1M tokens | Cached input per 1M | Output per 1M tokens |
gpt-rosalind-research | $5.00 | $0.50 | $25.00 |
gpt-rosalind-discovery | $5.00 | $0.50 | $25.00 |
gpt-6-astra (standard) | $10.00 | $1.00 | $50.00 |
........
OpenAI has not stated on its pricing page whether calls to gpt-6-rosalind are billed at the Rosalind Research rate or at GPT-6 Astra's rate, so organizations already in the program should confirm the billing that applies to the new model in their own console before moving production workloads to it. Billing for gpt-rosalind-research began on October 5, 2026, according to the same page. For broader context on OpenAI's current model pricing, see our coverage of GPT-6.1 Terra and its API pricing.
··········
WHAT IS ESTABLISHED AND WHAT REMAINS UNCLEAR.
The verified facts are clear enough. OpenAI released GPT-6 Rosalind on October 30, 2026 under the API identifier gpt-6-rosalind; it runs on the same weights as GPT-6 Astra without additional training; it substantially improves on GPT-Rosalind-5.5 on OpenAI's own medicinal chemistry, genomics and life sciences benchmarks; and it is deployed with a deliberately more permissive boundary on dual-use biology, offset by trusted access, separate safeguards, monitoring and mandatory human review. OpenAI classifies it as High, but not Critical, in biological and chemical capability, and as Critical in cybersecurity.
Several points remain open. All benchmark figures come from OpenAI, on internal or specialized test sets of around one hundred problems, and no independent evaluation has been published. The pricing of the new model is not yet listed as its own line, and the product surfaces beyond the API, such as Rosalind Workbench, ChatGPT and Codex, are not documented for this version in the materials OpenAI published on release day. The choice to ship a domain model by changing deployment policy rather than weights also raises a question for buyers: the capability gains come from the general flagship, so the value of Rosalind access lies mainly in fewer refusals on legitimate research questions and in the governance framework around them, which organizations will need to weigh against the obligations that come with approval. For related policy context, see our report on the UK AI Biosecurity Lab and its OpenAI and Anthropic credits.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
