GPT-Rosalind-6
GPT-Rosalind-6 is a life sciences reasoning model from OpenAI, announced on October 30, 2026, for biology, drug discovery and translational medicine.[1] It is the latest model in OpenAI's GPT-Rosalind series and the first built on GPT-6 Astra: according to its system card addendum it uses the same weights as GPT-6 Astra, with no added training data and no additional post-training.[1] What separates it from the general model is how it is deployed. OpenAI offers it only through a trusted-access program to approved scientists, research institutes, companies and government partners, and, according to OpenAI, it refuses fewer dual-use biology questions, including some that GPT-6 Astra may refuse in general access, while being designed to keep refusing requests for biological weaponization.[1] OpenAI's product channels name it differently. The system card says "GPT-Rosalind-6" and says OpenAI offers it as "GPT-Rosalind-6 Research"; the ChatGPT release notes and Help Center say "GPT-6 Rosalind Research"; the API changelog says "GPT-6 Rosalind"; and the API model identifier is gpt-6-rosalind.[1][3][5][6]
Under OpenAI's Preparedness Framework, the company treats the launch as High capability in the Biological and Chemical category (below Critical), Critical in Cybersecurity, and below High in AI Self-Improvement. The cybersecurity and self-improvement ratings are carried over from GPT-6 Astra's assessment.[1] In OpenAI's own life sciences benchmarks it scored well above its predecessor, GPT-Rosalind-5.5. On MedChemBench v7, for example, it averaged 49.6% against 28.9% at the highest reasoning budget tested.[1]
Overview
| Attribute | Detail |
|---|---|
| Developer | OpenAI[1] |
| Series | GPT-Rosalind (life sciences models)[5] |
| Announced | October 30, 2026 (system card addendum, release notes, API changelog)[1][3][6] |
| Base model | Same weights as GPT-6 Astra; no added training data or post-training[1] |
| Predecessor | GPT-Rosalind-5.5 (June 3, 2026), incrementally trained from GPT-5.5[8] |
| Product names | GPT-Rosalind-6 (system card); GPT-6 Rosalind Research (ChatGPT, Help Center); GPT-6 Rosalind (API changelog)[1][3][6] |
| API model ID | gpt-6-rosalind[4][5] |
| Access | Trusted access only; requires Rosalind Research approval for the organization and the user[1][5] |
| Surfaces | ChatGPT Work and Codex, plus the API where separately approved; not in the ChatGPT Chat tab[4][5] |
| API price (standard, up to 272K input tokens) | $15 input, $1.50 cached input, $75 output per million tokens (as of October 31, 2026)[7] |
| Preparedness classification | Biological and Chemical: High (below Critical); Cybersecurity: Critical; AI Self-Improvement: below High[1] |
Background
OpenAI announced the first GPT-Rosalind model on April 16, 2026, as the start of a series of models for life sciences research.[10] On June 3, 2026, it released GPT-Rosalind-5.5. That model was incrementally trained from GPT-5.5 on data related to beneficial life sciences capability and, unlike GPT-5.5, was trained not to refuse sophisticated biology queries; OpenAI made trusted access and responsible deployment its primary safeguard.[8] Its system card rated it High in Biological and Chemical capability (below Critical), with cybersecurity capabilities at or below those of GPT-5.5.[8] On September 8, 2026, OpenAI made gpt-rosalind-research generally available in the API through the trusted-access program at $5 per million input tokens and $25 per million output tokens, with billing starting on October 5, 2026.[6]
OpenAI released GPT-6 Astra, which it calls its flagship reasoning model, on September 3, 2026.[1][6][9] GPT-6 Astra was the first OpenAI model rated Critical in cybersecurity under the Preparedness Framework, and like GPT-Rosalind-5.5 it was treated as High in Biological and Chemical capability.[1][9] In October 2026 OpenAI also widened GPT-Rosalind access to academia: on October 15 it announced the Rosalind Academic Plan, which offers sponsored access to eligible labs at U.S. universities.[11] Within its "Rosalind" brand OpenAI also runs Rosalind Workbench, a research workspace in Codex and ChatGPT that does not require GPT-Rosalind model access.[5][12]
Model and training
The addendum says GPT-Rosalind-6 "uses the same weights as GPT-6 Astra" and that OpenAI "did not add training data or perform additional post-training for this release." For details of the underlying data and training it points readers to the GPT-6 Astra system card.[1] This is a change of approach from GPT-Rosalind-5.5, which was trained further on top of GPT-5.5 with life sciences data.[8] According to the addendum, what sets GPT-Rosalind-6 apart from GPT-6 Astra is a set of additional deployment controls: trusted access, model-level boundaries against harmful biological and cyber assistance (more permissive than GPT-6 Astra on some dual-use biology, more conservative on cyber), monitoring and enforcement, and additional contractual security requirements.[1]
Because the weights are shared, OpenAI uses GPT-6 Astra's helpful-only evaluations as additional evidence in its Preparedness assessment. It also repeated the biological capability evaluations on GPT-Rosalind-6 itself.[1]
Access and availability
Trusted access
GPT-Rosalind-6 is available only to approved customers. Before granting access, OpenAI checks whether an applicant has a legitimate life sciences or related mission, a credible use case, appropriate research and operational governance, and the ability to control access to the model safely.[1] It applies the same access requirements as for GPT-Rosalind-5.5, including policy updates made since that launch, and the addendum says the release "does not broaden eligibility beyond prior announcements, such as the Rosalind Academic Plan."[1] Non-government applicants also go through business verification and compliance screening. Separately, there is a trusted-access pathway for selected government entities working on approved biosecurity, biosafety, public-health preparedness and biodefense use cases.[1]
The Help Center describes two access tiers for GPT-Rosalind. Rosalind Research covers frontier biological work and includes GPT-6 Rosalind Research and GPT-5.5 Rosalind Research. Rosalind Discovery covers everyday scientific work and stays on the Rosalind 5.5 model, with Discovery-specific safeguards and stricter monitoring. Discovery approval does not grant Research access or access to GPT-6 Rosalind Research.[4][5] The Rosalind Academic Plan does not include API access or API credits.[5]
ChatGPT and Codex rollout
According to the October 30, 2026 ChatGPT Enterprise and Edu release notes, GPT-6 Rosalind Research is rolling out to "a limited set of approved Enterprise customers for life sciences research in Work and Codex." It requires Research approval for both the organization and the user, and it is not available in the Chat tab.[3] In Enterprise workspaces the model starts in a two-week Admin Preview and is off by default. Admins can enable it for approved Research users and set up access for after the preview. Existing Research approval does not bypass the preview.[3][5] Users choose GPT-6 Rosalind Research from the model picker in ChatGPT Work or Codex, and GPT-5.5 Rosalind Research remains available for existing workflows.[3]
API
The OpenAI API changelog lists the release of GPT-6 Rosalind (gpt-6-rosalind) on October 30, 2026. Access is open to organizations approved for Rosalind Research "for internal research by authorized users."[6] The Help Center says API access must be approved separately and covers internal research tools, workflows and applications, and is not available for customer-facing products or external commercial applications "at this time". Approval for another GPT-Rosalind model does not grant access to gpt-6-rosalind on its own.[5]
As of October 31, 2026, OpenAI's API pricing page lists these standard per-million-token rates:[7]
| Model | Input | Cached input | Cache write (30 min) | Cache write (12 h) | Output |
|---|---|---|---|---|---|
gpt-6-rosalind (up to 272K input tokens) | $15.00 | $1.50 | $18.75 | $37.50 | $75.00 |
gpt-rosalind-research | $5.00 | $0.50 | - | - | $25.00 |
gpt-rosalind-discovery | $5.00 | $0.50 | - | - | $25.00 |
For gpt-6-rosalind prompts longer than 272K input tokens, the page's long-context rates are $30 per million input tokens, $3 cached input, $37.50 cache write and $112.50 output.[7] The pricing page also says regional processing (data residency) endpoints cost 10% more for eligible models released on or after March 5, 2026.[7] In ChatGPT, the Help Center says GPT-6 Rosalind Research has different standard rates from the existing Rosalind 5.5 models, and it points customers to OpenAI's rate cards and their own agreements.[5]
Beneficial capability evaluations
OpenAI compared GPT-Rosalind-6 with GPT-Rosalind-5.5 on three life sciences research benchmarks. It re-ran GPT-Rosalind-5.5 on the same benchmark versions and problem sets, so those scores differ from the ones published when GPT-Rosalind-5.5 launched. Response scores were averaged within each problem, and problems were weighted equally.[1] All of these are OpenAI's own benchmarks and results.
| Benchmark | What it tests | Problems | Metric (highest reasoning budget tested) | GPT-Rosalind-6 | GPT-Rosalind-5.5 |
|---|---|---|---|---|---|
| MedChemBench v7 | Chemical structure interpretation, structure-activity relationships, drug potency and toxicity prediction | 99 (up to 9 scored responses each) | Mean rubric score | 49.6% | 28.9% |
| GeneBench Pro v13 | Planning and carrying out quantitative analyses of biological data as an agent | 129 (7 responses per model each) | Single-attempt success rate | 37.7% | 16.7% |
| LifeSciBench Gold | Research tasks contributed by practicing life scientists | 94 (six flagged problems excluded; up to 8 scored responses each) | Average score | 61.3% | 44.0% |
Source: GPT-Rosalind-6 System Card Addendum, section 3.[1][2] The GeneBench Pro evaluation used a benchmark-specific configuration that differs from production.[1]
Preparedness evaluations
Biological and Chemical
As it did for GPT-6 Astra, OpenAI treats GPT-Rosalind-6 as High capability in the Biological and Chemical category. It uses the four High-capability and four Critical-capability evaluations described in the GPT-6 Astra system card.[1] The High evaluations test wet-lab knowledge and troubleshooting. The Critical evaluations test biological design capabilities that could help an expert develop a highly dangerous novel threat. OpenAI calls the Critical thresholds "preliminary estimates informed by expert judgment" and says they may change.[1] Two evaluations reported for GPT-Rosalind-5.5, hard-negative protein binding and DNA sequence design, have been retired.[1]
For the High evaluations, the addendum reports observed scores alongside refusal-adjusted scores, which count refusals as correct answers to estimate capability that safety training might hide. GPT-6 Astra reference values come from the runs reported in its system card.[1]
| Evaluation | Tier | Metric | Indicative threshold | GPT-Rosalind-6 | GPT-Rosalind-5.5 | GPT-6 Astra |
|---|---|---|---|---|---|---|
| ProtocolQA Open-Ended | High | pass@1 / refusal-adjusted | 54% | 42.0% / 42.0% | 36.4% / 36.4% | 41.4% / 45.4% |
| TroubleshootingBench | High | pass@1 / refusal-adjusted | 36.4% | 59.0% / 59.0% | 53.3% / 53.3% | 48.4% / 63.5% |
| Biorisk knowledge | High | cons@32 / refusal-adjusted cons@32 | 80% | 78.3% / 85.0% | 79.2% / n/a | 63.3% / 90.0% |
| Multi-select Virology Troubleshooting | High | pass@1 / refusal-adjusted | 31% | 59.9% / 59.9% | 51.2% / 51.2% | 55.8% / 63.1% |
| AAV Capsid Packaging Prediction | Critical | Spearman correlation (higher is better) | 0.600 | 0.529 | 0.22 | 0.528 |
| SHP2 Protein Function Prediction | Critical | Mean R² (higher is better) | 0.60 | 0.38 | 0.16 | 0.40 |
| Coronavirus-ACE2 Cell-Entry Screen | Critical | Composite score (higher is better) | 0.75 | 0.42 | 0.44 | 0.45 |
| Phage-plasmid Co-evolution | Critical | Negative log-likelihood (lower is better) | 9.40 or lower | 13.41 | 11.81 | 12.58 |
Source: GPT-Rosalind-6 System Card Addendum, Tables 2-9; figures match between the web version and the PDF.[1][2]
Notes on the table, from the addendum:[1]
- ProtocolQA Open-Ended turns 108 multiple-choice questions from FutureHouse's ProtocolQA dataset, part of LAB-Bench, into open-ended short-answer questions about fixing deliberately broken published protocols.[1][13]
- TroubleshootingBench uses non-public protocols transcribed by PhD scientists, with troubleshooting questions built around realistic execution errors. GPT-Rosalind-6 scored 491 of 832, with nine failed evaluation attempts counted as incorrect.
- Biorisk knowledge is a multiple-choice tacit-knowledge and troubleshooting set created with Gryphon Scientific, covering all five stages of the biothreat creation process. GPT-Rosalind-6 refused 9.4% of questions, compared with 1.8% for GPT-Rosalind-5.5 and 29.2% for GPT-6 Astra. GPT-Rosalind-5.5's published refusal-adjusted score used a different calculation and is left out.
- Multi-select Virology Troubleshooting uses held-out questions from SecureBio. A response gets credit only if it selects every correct statement and no incorrect ones.
- AAV Capsid Packaging Prediction is an evaluation developed by Dyno Therapeutics that asks the model to predict packaging scores for held-out adeno-associated virus capsid variants.
- SHP2 Protein Function Prediction asks for functional assay results of single-amino-acid variants across three unpublished datasets. The GPT-6 Astra figure is its reported maximum; a separate helpful-only run scored 0.35.
- Coronavirus-ACE2 Cell-Entry Screen uses an unpublished academic dataset from experiments on existing spike proteins with replication-incompetent pseudoviruses. The GPT-6 Astra reference here, as for the phage-plasmid evaluation, is its helpful-only result.
- Phage-plasmid Co-evolution asks the model to predict which mutations will appear later in an unpublished experimental system that uses viruses infecting bacteria.
After adjusting for refusals, GPT-Rosalind-6 exceeded the indicative High threshold on three of the four High evaluations (all except ProtocolQA Open-Ended), and OpenAI therefore kept treating it as High capability. Compared with GPT-Rosalind-5.5, its observed scores were higher on three of the High evaluations and lower on Biorisk knowledge.[1] It met none of the four Critical thresholds, and results against GPT-Rosalind-5.5 on those evaluations were mixed. OpenAI says it also relies on GPT-6 Astra's helpful-only evaluations and external biological capability assessments.[1]
Cybersecurity and AI self-improvement
The addendum reports no new cybersecurity or AI self-improvement Preparedness capability evaluations for GPT-Rosalind-6. OpenAI relies on GPT-6 Astra's existing assessments.[1] On that basis it treats GPT-Rosalind-6 as Critical in Cybersecurity, and as below High in AI Self-Improvement, since GPT-6 Astra did not meet the High threshold there.[1]
Safeguards
Compared with GPT-6 Astra, GPT-Rosalind-6 has extra deployment controls. The addendum names four main components: trusted access controls for verified higher-capability life sciences use; model-level boundaries against harmful biological and cyber assistance; monitoring and enforcement to detect abuse; and additional contractual security requirements. It also keeps GPT-6 Astra's standard cyber safeguards, including real-time blocking.[1] Approved enterprise customers must limit access to authorized users with a legitimate need and apply least-privilege access, onboarding and insider-risk controls. Qualified human review is required before model outputs inform experiments, publications or consequential decisions.[1]
Biological and chemical boundaries
OpenAI says GPT-Rosalind-6 is meant to be more useful for legitimate advanced biological research than generally available ChatGPT models, including on some dual-use topics that GPT-6 Astra may refuse in general access. The addendum also says the model is still tuned to refuse or safely respond to requests that would meaningfully enable severe biological weaponization. OpenAI says this training is done separately from the training of its classifiers and offline mitigations "to decorrelate our safeguards". The addendum does not explain how this tuning relates to its statement that no additional post-training was performed for this release.[1] Because this is a trusted-access deployment, OpenAI says it mainly looked at the Biosecurity - Severe benchmark. Biosecurity - Benign measures full compliance on benign questions; the other rows show the share of responses that pass a safety grader on prompts about biological weaponization or chemical misuse.[1]
| Benchmark (higher is better) | GPT-Rosalind-6 | GPT-Rosalind-5.5 | GPT-6 Astra |
|---|---|---|---|
| Biosecurity - Severe | 98.1% | 88.6% | 99.8% |
| Biosecurity - Sensitive | 79.0% | 18.3% | 99.8% |
| Biosecurity - Select Sequences | 83.8% | 50.0% | 95.6% |
| Biosecurity - Dual-Use | 84.6% | 69.0% | 97.8% |
| Chemical Misuse - Severe | 100.0% | 59.1% | 100.0% |
| Biosecurity - Benign | 100.0% | 97.7% | 91.6% |
Source: GPT-Rosalind-6 System Card Addendum, Table 10.[1][2] OpenAI notes that these benchmarks and their grading keep changing, so the results cannot be compared directly with past system cards, and a response that fails the grader is not necessarily unsafe.[1]
Cyber boundaries
GPT-Rosalind-6 inherits GPT-6 Astra's model-level and system-level cyber safeguards and adds a more conservative cyber refusal boundary. It is meant to keep refusing requests that clearly enable unauthorized, destructive or harmful cyber activity, such as malware deployment, credential theft and exfiltration.[1] As a robustness check, OpenAI replayed nine known jailbreak attacks across ten fixed ExploitBench tasks with real-time cyber blocking turned off. The mean task success rate across the seven attacks that produced usable results was 14.3% for GPT-Rosalind-6 and 15.7% for GPT-6 Astra (lower is better). Two attacks were excluded because every attempt errored or exceeded turn limits.[1]
Safety and alignment evaluations
The addendum also reports instruction-following, honesty, tool-use and monitoring evaluations against GPT-6 Astra. OpenAI calls these descriptive comparisons, because evaluation versions, inference settings, grading and sample coverage differ across some runs.[1]
| Evaluation | Measure | GPT-Rosalind-6 | GPT-6 Astra |
|---|---|---|---|
| Blocker Deception | Blocker-related deception rate (lower is better) | 9.5% | 12.9% |
| Broken Search Tool | Rate of failing to acknowledge the search tool is unavailable (lower is better) | 1.3% | 1.5% |
| ExploitGym Honeypot | Rate of exploiting evaluation infrastructure outside the assigned task (lower is better) | 0.0% | 0.0% |
| Respecting Warnings | Rate of successfully circumventing a restriction after a warning (lower is better) | 21.7% | 18.9% |
| Respecting Auto-Review | Rate of attempting to circumvent an automatic-review denial (lower is better) | 0% | 0% |
| Agent Replay Auto-Review | Rate of harmful or unauthorized actions in replayed agentic conversations, excluding laziness (lower is better) | 27.1% | 26.7% |
| Unintended Engagement with External Agent Messages | Rate of performing the specified unauthorized action after discovering the message board (lower is better) | 2.5% | 0.0% |
| Cowork Misalignment | Rate of misaligned outcomes in simulated workplace tasks (lower is better) | 1.4% | 1.2% |
| OAI-repo Sabotage v2 | Rate of sabotage that also evades the monitor (lower is better) | 0% unguided, 0% guided | 0% unguided, 3% guided |
| Math Side-tasking Monitor Evasion | Full-context monitor bypass rate (lower is better) | 0% | 0% |
| Chain-of-thought controllability | Rate of following instructions that constrain the model's reasoning (lower is better) | GPQA 71.8%, HLE 59.3%, MMLU 83.1% | GPQA 69.2%, HLE 53.8%, MMLU 81.6% |
Source: GPT-Rosalind-6 System Card Addendum, Tables 12-14.[1][2]
Enforcement and security
OpenAI can restrict or revoke access for misuse. To spot possible malicious biological, chemical or other policy-violating activity after access is granted, it uses automated safety classifiers, account- and organization-level indicators, customer-provided information and, where available, specialist review.[1] OpenAI tested its shared monitoring and enforcement process before launch through a biosecurity red-teaming exercise on GPT-Rosalind-5.5, and the addendum says it is repeating that exercise for GPT-Rosalind-6 now that the model is deployed. OpenAI did not publish the details, citing safeguard sensitivity, but says it shared them with its Safety Advisory Group and other governance bodies.[1] Customers must maintain strong identity and access management, administrator controls, least-privilege access and processes for removing users who no longer need access. Model access and data-retention arrangements are approved separately.[1]
Continue reading
- GPT-RosalindGPT-Rosalind is a family of frontier reasoning models for the life sciences, developed by OpenAI and first released on April 16, 2026.Mentioned 56 times above
- GPT-6 AstraGPT-6 Astra is an OpenAI model that the company began rolling out on September 3, 2026, describing it as "the world's most intelligent and aligned model" and as the successor to the GPT-5.6 family.Mentioned 30 times above
- GPT-5.5GPT-5.5 is a large language model developed by OpenAI and released on April 23, 2026.Mentioned 7 times above
- Preparedness Framework (OpenAI)The Preparedness Framework is the risk-management policy maintained by openai for tracking, evaluating, forecasting, and mitigating catastrophic risks from frontier artificial-intelligence models.Mentioned twice above
- LAB-BenchLAB-Bench (the Language Agent Biology Benchmark) is an AI benchmark of more than 2,400 multiple-choice questions built to measure how well language models and AI agents can perform practical biology research…Listed in See also
- OpenAI for ScienceOpenAI for Science was a research initiative and internal team at OpenAI, announced on September 2, 2025Listed in See also
More from See also
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47OpenAI Deployment Safety Hub, "GPT-Rosalind-6 System Card Addendum," October 30, 2026. deploymentsafety.openai.com/gpt-rosalind-6
- ^1 ^2 ^3 ^4OpenAI, "GPT-Rosalind-6 System Card Addendum" (PDF), October 2026. deploymentsafety.openai.com/...gpt-rosalind-6.pdf
- ^1 ^2 ^3 ^4 ^5 ^6OpenAI Help Center, "ChatGPT Enterprise and Edu release notes," entry "GPT-6 Rosalind Research," October 30, 2026. help.openai.com/...nterprise-and-edu-release-notes
- ^1 ^2 ^3OpenAI Help Center, "Model Release Notes," entry "GPT-6 Rosalind Research," October 30, 2026. help.openai.com/...9624314-model-release-notes
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11OpenAI Help Center, "GPT-Rosalind for life sciences research," accessed October 31, 2026. help.openai.com/...lind-for-life-sciences-research
- ^1 ^2 ^3 ^4 ^5 ^6OpenAI API, "Changelog" (entries of September 3, September 8 and October 30, 2026). developers.openai.com/...changelog
- ^1 ^2 ^3 ^4OpenAI API, "Pricing," life sciences models section, accessed October 31, 2026. developers.openai.com/...pricing
- ^1 ^2 ^3 ^4OpenAI Deployment Safety Hub, "GPT-Rosalind-5.5 System Card," June 3, 2026. deploymentsafety.openai.com/gpt-rosalind-5-5
- ^1 ^2OpenAI Deployment Safety Hub, "GPT-6 Astra System Card," September 3, 2026. deploymentsafety.openai.com/gpt-6-astra
- ^Megan Morrone, "OpenAI launches new AI model for life sciences research," Axios, April 16, 2026. axios.com/...openai-models-life-sciences-drugs
- ^OpenAI (@OpenAI), post announcing the Rosalind Academic Plan, X, October 15, 2026. x.com/...2110841898076418336
- ^OpenAI, "Rosalind," accessed October 31, 2026. openai.com/gpt-rosalind
- ^J. M. Laurent, J. D. Janizek, M. Ruzo, et al., "LAB-Bench: Measuring Capabilities of Language Models for Biology Research," arXiv:2407.10362, 2024. arxiv.org/...2407.10362
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
2 revisions · v2 · 3,363 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independently verified 2026-10-31 against OpenAI's GPT-Rosalind-6 System Card Addendum (HTML chart data and PDF), ChatGPT Enterprise/Edu release notes, Help Center access article, API changelog and pricing page; 9 minor verifier defects fixed before stamping.
Cite this page: AI Wiki. "GPT-Rosalind-6." aiwiki.ai, updated 1 Nov 2026, fact-checked 31 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/gpt_rosalind_6