{"slug": "gpt-rosalind-6-system-card-addendum", "title": "GPT-Rosalind-6 System Card Addendum", "summary": "OpenAI released a system card addendum for GPT-Rosalind-6, a frontier biology and drug-discovery reasoning model built on the same weights as its flagship GPT-6 Astra with no additional training data or post-training. OpenAI classifies the launch as High capability in the Biological and Chemical domain under its Preparedness Framework, below the Critical threshold, while treating GPT-Rosalind-6 as Critical in Cybersecurity and below High in AI Self-Improvement. The model, offered as GPT-Rosalind-6 Research to vetted scientists, research institutes, companies and government partners, scored 49.6% versus 28.9% for GPT-Rosalind-5.5 on MedChemBench v7, 37.7% versus 16.7% single-attempt success on GeneBench Pro v13, and 61.3% versus 44.0% on LifeSciBench Gold at the highest reasoning budget tested.", "body_md": "GPT-Rosalind-6 is a frontier reasoning model for biology, drug\ndiscovery, and translational medicine. We offer it as GPT-Rosalind-6\nResearch to qualified scientists, research institutes, companies and\ngovernment partners. Rosalind has fewer refusals on dual-use biological\nqueries but is only made available via trusted access. We review\napplicants’ proposed uses, security controls and governance practices\nbefore granting access. We also continue to have other deployment\nsafeguards in place.\n\nGPT-Rosalind-6 uses the same weights as GPT-6 Astra, our flagship\nreasoning model. It is designed to support legitimate biological\nresearch, including some dual-use topics, while still refusing requests\nfor biological weaponization and retaining refusals on other safety\ntopics.\n\nUnder our Preparedness Framework, we are treating this launch as High\ncapability in the Biological and Chemical domain, below the threshold\nfor Critical. We rely on GPT-6 Astra’s capability assessments for other\nPreparedness risks and treat GPT-Rosalind-6 as Critical in Cybersecurity\nand below High in AI Self-Improvement.\n\nThis addendum describes GPT-Rosalind-6’s biological capability\nassessment, additional safety and alignment evaluations, and deployment\nsafeguards. We include results for GPT-Rosalind-5.5 where available for\ncomparison. This addendum should be read alongside the GPT-6 Astra\nsystem card, which describes the parent model’s broader capability\nand safety evaluations.\n\n2 Model Data and Training\n\nGPT-Rosalind-6 uses the same weights as GPT-6 Astra. We did not add\ntraining data or perform additional post-training for this release.\nDetails of the underlying model’s data and training are provided in the\nGPT-6 Astra\nsystem card.1\n\n3 Beneficial Capability Evaluations\n\nWe evaluated GPT-Rosalind-6 on three life sciences research\nbenchmarks and compared its results with GPT-Rosalind-5.5. The\nGPT-Rosalind-5.5 scores below use the same benchmark versions and\nproblem sets as GPT-Rosalind-6 and therefore differ from those reported\nat its launch. The Preparedness evaluations in Section 4 assess\ncapabilities relevant to the risk of severe harm. We averaged response\nscores within each problem and then weighted the problems equally.\n\n3.1 Medicinal chemistry\n\nMedChemBench v7 tests chemical structure interpretation,\nstructure-activity relationships, and predictions of drug potency and\ntoxicity. We evaluated both models on the same 99 problems, with up to\nnine scored responses per problem at each reasoning budget. At the\nhighest reasoning budget tested, mean rubric scores were 49.6% for\nGPT-Rosalind-6 and 28.9% for GPT-Rosalind-5.5.\n\n3.2 Genomics and quantitative biology\n\nGeneBench Pro v13 tests an agent’s ability to plan and carry out\nquantitative analyses of biological data. We estimated single-attempt\nsuccess from seven responses per model on each of 129 problems at each\nreasoning budget. At the highest reasoning budget tested, single-attempt\nsuccess rates were 37.7% for GPT-Rosalind-6 and 16.7% for\nGPT-Rosalind-5.5. The evaluation used a benchmark-specific configuration\nthat differs from production.\n\n3.3 Applied life sciences research\n\nLifeSciBench Gold uses research tasks contributed by practicing life\nscientists. We excluded six problems flagged for possible issues with\nthe tasks or evaluation environment and compared both models on the same\nremaining 94 problems, with up to eight scored responses per problem at\neach reasoning budget. At the highest reasoning budget tested, average\nscores were 61.3% for GPT-Rosalind-6 and 44.0% for GPT-Rosalind-5.5.\n\n4 Preparedness\n\nThe Preparedness Framework is OpenAI’s approach to tracking and\npreparing for frontier capabilities that create new risks of severe\nharm. Under our framework, we work to track and mitigate the risk of\nsevere harm, including by implementing safeguards that sufficiently\nminimize the risk for highly capable models.\n\n4.1 Preparedness Evaluations\n\nWe repeated biological capability evaluations on GPT-Rosalind-6 and\nuse the parent model’s helpful-only evaluations as additional evidence,\ngiven it uses the same weights as GPT-6 Astra. Helpful-only models are\nevaluated to elicit capabilities that safety refusals may otherwise\nconceal.\n\n4.1.1 Biological and Chemical\n\nAs we did for GPT-6 Astra, we are treating this launch as High\ncapability in the Biological and Chemical domain. Below, we describe the\nevaluations we used to reach this conclusion. We use the four High\ncapability evaluations and four Critical capability evaluations\ndescribed in the GPT-6 Astra system card. The hard-negative protein\nbinding and DNA sequence design evaluations reported for\nGPT-Rosalind-5.5 have been retired.\n\nThe High capability evaluations test wet lab knowledge and\ntroubleshooting. The Critical capability evaluations test biological\ndesign capabilities that could help an expert develop a highly dangerous\nnovel threat. These evaluations provide evidence about specific\ncapabilities rather than a complete test of the ability to cause\nbiological harm. The Critical capability thresholds are preliminary\nestimates informed by expert judgment and may change as the evaluations\nare refined.\n\nTable 1: Overview of Biological and Chemical evaluations\n\nTable 1: Overview of Biological and Chemical evaluations\n\nEvaluation\n\nCapability\n\nDescription\n\nHigh\n\nProtocolQA Open-Ended\n\nWet lab capabilities\n\nTroubleshooting commonly published lab protocols.\n\nTroubleshootingBench\n\nTacit knowledge and troubleshooting\n\nIdentifying errors in non-public, expert-written lab protocols.\n\nBiorisk knowledge\n\nBiorisk-relevant knowledge\n\nAnswering questions about tacit biological knowledge and laboratory troubleshooting.\n\nMulti-select Virology Troubleshooting\n\nWet lab troubleshooting\n\nIdentifying the complete set of correct responses to virology troubleshooting questions.\n\nCritical\n\nAAV Capsid Packaging Prediction\n\nNovel Design (Protein design)\n\nPredicting packaging performance for held-out capsid variants.\n\nSHP2 Protein Function Prediction\n\nNovel Design (Protein design)\n\nPredicting functional assay results for protein variants.\n\nCoronavirus-ACE2 Cell-Entry Screen\n\nNovel Design (Virus-host interaction)\n\nPredicting and ranking held-out cell-entry measurements.\n\nPhage-plasmid Co-evolution\n\nNovel Design (Co-evolution)\n\nPredicting mutations that appear at later experimental timepoints.\n\nThe results below report observed performance and, for the High\ncapability evaluations, refusal-adjusted scores. pass@1 measures success\non a single attempt. The adjustment treats refusals as correct answers\nto estimate capability that safety training may otherwise conceal. For\nBiorisk knowledge, we recompute the consensus answer after this\nadjustment. GPT-6 Astra reference values use the evaluation runs\nreported in its system card.\n\n4.1.1.1 ProtocolQA Open-Ended\n\nTo evaluate models’ ability to troubleshoot commonly published lab\nprotocols, we modify 108 multiple-choice questions from FutureHouse’s\nProtocolQA dataset [3]\nto be open-ended short answer questions, which makes the evaluation\nharder and more realistic than the multiple-choice version. The\nquestions introduce egregious errors in common published protocols,\ndescribe the wet lab result of carrying out this protocol, and ask how\nto fix the procedure.\n\nTable 2: ProtocolQA Open-Ended. The indicative High capability threshold is 54%.\n\nCategory\n\n4.1.1.2 TroubleshootingBench\n\nTo evaluate models’ ability to identify and correct real-world\nexperimental errors in biological protocols, we built a short-answer\ntroubleshooting dataset from expert-written wet lab procedures.\nTroubleshootingBench focuses on tacit, hands-on knowledge and\nuncontaminated procedures that are not available online.\n\nScientists with a PhD in a relevant biological discipline were asked\nto transcribe biological protocols they have personally used in the lab.\nFrom these protocols, they created troubleshooting questions,\nintroducing subtle or realistic execution errors and describing the\nresulting failed outcome. Compared to ProtocolQA Open-Ended, which\nfocuses on well-known published procedures, TroubleshootingBench tests\nperformance on non-public protocols and errors that rely on tacit\nprocedural knowledge.\n\nTable 3: TroubleshootingBench. The indicative High capability threshold is 36.4%. GPT-Rosalind-6 scored 491/832, counting nine failed evaluation attempts as incorrect.\n\nCategory\n\n4.1.1.3 Biorisk knowledge\n\nWe evaluated models on a tacit knowledge and troubleshooting\nmultiple-choice dataset created with Gryphon Scientific. The questions\nspan all five stages in the biothreat creation process and focus on\nareas where tacit knowledge would be a bottleneck. Tacit knowledge\nquestions are meant to be obscure to anyone not working in the field.\nTroubleshooting questions are meant to be obscure to anyone without\nhands-on experience.\n\nTable 4: Biorisk knowledge. cons@32 uses the consensus answer across 32 responses to each question. The indicative High capability threshold is 80%.\n\nTable 4: Biorisk knowledge. cons@32 uses the consensus answer across 32 responses to each question. The indicative High capability threshold is 80%.\n\nMetric\n\nGPT-Rosalind-6\n\nGPT-Rosalind-5.5\n\nGPT-6 Astra\n\ncons@32\n\n78.3%\n\n79.2%\n\n63.3%\n\nRefusal rate\n\n9.4%\n\n1.8%\n\n29.2%\n\nRefusal-adjusted cons@32\n\n85.0%\n\nn/a\n\n90.0%\n\nThe refusal-adjusted GPT-Rosalind-5.5 score published in its system\ncard used a different calculation and is therefore omitted from this\ncomparison.\n\n4.1.1.4 Multi-select Virology Troubleshooting\n\nTo evaluate models’ ability to troubleshoot wet lab experiments in a\nmultimodal setting, we evaluate models on a set of fully held-out\nvirology troubleshooting questions from SecureBio, in a multi-select\nformat. A response receives credit only if it selects every correct\nstatement and no incorrect ones.\n\nTable 5: Multi-select Virology Troubleshooting. The indicative High capability threshold is 31%.\n\nCategory\n\nAfter adjusting for refusals, GPT-Rosalind-6 exceeds the indicative\nHigh capability thresholds on three of the four evaluations. We\ntherefore continue to treat the model as High capability. Compared with\nthe reported GPT-Rosalind-5.5 results, observed scores are higher on\nthree evaluations and lower on Biorisk knowledge.\n\n4.1.1.5 AAV Capsid Packaging Prediction\n\nPredicting whether an AAV capsid variant will package successfully\ncould reduce trial and error in viral-vector engineering, which we treat\nas a proxy for some dual-use biological design capabilities. We evaluate\nthis capability using an evaluation developed by Dyno Therapeutics. The\nmodel predicts packaging scores for held-out sequences using labeled\nexamples, protein embeddings and computational tools. Performance is\nmeasured using Spearman rank correlation.\n\nTable 6: AAV Capsid Packaging Prediction. The indicative Critical capability threshold is 0.600.\n\nCategory\n\n4.1.1.6 SHP2 Protein Function Prediction\n\nPredicting how changes to protein sequence affect function could\nreduce trial and error in protein engineering, which we treat as a proxy\nfor some dual-use biological design capabilities. We evaluate this\ncapability using SHP2, a well-studied human protein. The model uses\navailable experimental measurements to predict held-out functional assay\nresults for single-amino-acid variants. Performance is measured by mean\nR² across three unpublished assay datasets.\n\nTable 7: SHP2 Protein Function Prediction. The indicative Critical capability threshold is 0.60. The GPT-6 Astra reference is the reported maximum of 0.40. A separate helpful-only run scored 0.35.\n\nCategory\n\n4.1.1.7 Coronavirus-ACE2 Cell-Entry Screen\n\nReliable predictions of cell entry could help researchers prioritize\nlaboratory experiments in studies of virus-host interactions, but could\nalso be a threat-relevant capability. We ask models to predict and rank\nheld-out measurements from an existing, unpublished academic dataset.\nThe underlying experiment studied existing spike proteins using\nreplication-incompetent pseudoviruses. It did not introduce new\nmutations or generate infectious progeny.\n\nTable 8: Coronavirus-ACE2 Cell-Entry Screen. The indicative Critical capability threshold is 0.75. The GPT-6 Astra reference is its helpful-only result.\n\nCategory\n\n4.1.1.8 Phage-plasmid Co-evolution\n\nWe assess the ability to predict biological evolution using data from\nan existing but unpublished experimental phage-plasmid co-evolution\nsystem. Given observations available at each experimental checkpoint,\nthe model predicts which mutations will subsequently appear. Performance\nis measured by the negative log-likelihood of observed mutations, with\nlower scores indicating better predictions. The experiment uses viruses\nthat infect bacteria rather than humans. We treat performance as a proxy\nfor some dual-use biological design capabilities.\n\nTable 9: Phage-plasmid Co-evolution. The indicative Critical capability threshold is an NLL of 9.40 or lower. The GPT-6 Astra reference is its helpful-only result.\n\nCategory\n\nNone of the four Critical capability evaluations met its indicative\nthreshold. Results relative to GPT-Rosalind-5.5 were mixed. We also rely\non the parent model’s helpful-only evaluations and external biological\ncapability assessments, as described in the GPT-6 Astra system card.\n\n4.1.2 Cybersecurity\n\nWe rely on GPT-6 Astra’s existing Preparedness assessment for\nCybersecurity and treat GPT-Rosalind-6 as Critical in this domain. The\nunderlying capability findings are described in the GPT-6 Astra\nsystem card.\n\n4.1.3 AI Self-Improvement\n\nWe rely on GPT-6 Astra’s existing Preparedness assessment for AI\nSelf-Improvement, where GPT-6 Astra did not meet the threshold for High\ncapability. The evaluations and findings are described in the GPT-6 Astra\nsystem card.\n\n4.2 Safeguards\n\nCompared to GPT-6 Astra, GPT-Rosalind-6 features additional\ndeployment controls for its use in advanced life sciences research.\nAccess is limited to approved customers through our trusted-access\ndeployment structure where organizations must demonstrate they are\nconducting legitimate scientific research with public benefit, have\nstrong governance and safety oversight, and control access with\nenterprise-grade security.\n\nOur safeguard approach for GPT-Rosalind-6 has four main components:\ntrusted access controls for verified higher-capability life sciences\nuse, model-level boundaries against harmful biological and cyber\nassistance, monitoring and enforcement mechanisms designed to detect\nabuse, and additional contractual requirements to enhance security.\n\nGPT-Rosalind-6 retains GPT-6 Astra’s standard cyber safeguards,\nincluding real-time blocking.\n\n4.2.1 Trust-based access\n\nGPT-Rosalind-6 is available only to approved customers. Before\ngranting access, OpenAI reviews whether the applicant has a legitimate\nlife sciences or related mission, a credible use case, appropriate\nresearch and operational governance, and the ability to safely control\naccess to the model.\n\nGPT-Rosalind-6 uses the access requirements already applied to\nGPT-Rosalind-5.5, including the policy updates introduced since its\nlaunch. This release does not broaden eligibility beyond prior\nannouncements, such as the Rosalind Academic Plan.\n\nFor non-government applicants, OpenAI also employs business\nverification and compliance-screening checks. These checks help confirm\napplicant authenticity and reduce the risk of access by fraudulent,\ninsufficiently attributable, or otherwise unsuitable actors. These\nchecks supplement OpenAI’s broader review of the applicant’s proposed\nuse case, safety posture, and access scope.\n\nFor government applicants, OpenAI maintains a trusted-access pathway\nfor select government entities working on approved biosecurity,\nbiosafety, public-health preparedness, and biodefense use cases.\n\nApproved enterprise customers must limit GPT-Rosalind-6 access to\nauthorized users with a legitimate need and apply least-privilege access\ncontrols, appropriate onboarding, and insider risk controls to their\nusers. Customers must apply qualified human review before model outputs\ninform experiments, publications, or consequential decisions.\n\n4.2.2 Model behavior and system safeguards\n\nThese evaluations were deliberately created to be difficult. They\nwere built around cases in which our existing models were not yet giving\nideal responses, and this is reflected in the scores below. We continue\nmonitoring these categories after launch to evaluate online performance\nand further adjust safeguards as appropriate.\n\n4.2.2.1 Biological and Chemical safeguards\n\nGPT-Rosalind-6 is designed to be more useful for legitimate advanced\nbiological research than generally available ChatGPT models. This\nsupports advanced life sciences workflows, including some dual-use\ntopics, that GPT-6 Astra may refuse in General Access. We still tune the\nmodel to refuse or safely respond to requests that would continue to\nmeaningfully enable severe biological weaponization. This training is\ndone separately from the training of our classifiers and offline\nmitigations to decorrelate our safeguards.\n\nTo evaluate the quality of these model-level refusals, we track the\nsafety of model responses to prompts that originate from held-out\nsynthetic data, red-teaming, and production data. These benchmarks and\ntheir grading continue to be updated, meaning results do not directly\ntranslate to past system cards.\n\nTable 10 reports\nresults across our suite of biosecurity benchmarks, accounting for both\nmodel behavior and safeguards. We note that because GPT-Rosalind-6 is a\ntrusted-access deployment, we mainly considered the results from the\nBiosecurity - Severe benchmark. Biosecurity - Benign\nevaluates model compliance (being fully helpful without refusals or\ncaveats) on a set of benign biological and operational questions. The\nremaining benchmark scores show the fraction of model responses that\npass the configured safety grader on prompts related to biological\nweaponization. A response that does not pass was flagged under this\ngrading criterion; this does not necessarily mean it is unequivocally\nunsafe.\n\nTable 10: Biological and chemical content evaluations (higher is better)\n\nCategory\n\n4.2.2.2 Cyber boundaries\n\nGPT-Rosalind-6 inherits the model and system level cyber safeguards\nof the GPT-6 Astra base model and additionally applies a more\nconservative cyber refusal boundary. The model is intended to continue\nto refuse requests that clearly enable unauthorized, destructive, or\nharmful cyber activity, including areas such as malware deployment,\ncredential theft, exfiltration, or other malicious activity.\n\nAs an additional robustness test, we evaluate resistance to\npreviously identified cyber jailbreaks by replaying nine known attacks\nacross ten fixed ExploitBench tasks. We evaluated both models with\nreal-time cyber blocking disabled, so the task success rates in\nTable 11 reflect\nmodel-level behavior only. We report the mean task success rate across\nseven attacks, excluding attempts that ended in errors or exceeded turn\nlimits.\n\nTable 11: Task success rates on historical cyber jailbreak attacks with real-time blocking disabled (lower is better).\n\nTable 11: Task success rates on historical cyber jailbreak attacks with real-time blocking disabled (lower is better).\n\nEvaluation\n\nMeasure\n\nGPT-Rosalind-6\n\nGPT-6 Astra\n\nHistorical Jailbreak Regression\n\nTask success rate under jailbreak attacks ↓\n\n14.3%\n\n15.7%\n\nTwo attacks were excluded because all attempts ended in errors or\nexceeded turn limits.\n\n4.2.2.3 Safety and alignment evaluations\n\nWe also evaluate GPT-Rosalind-6’s instruction following, honesty,\nresistance to prompt injection, and behavior when using tools.\nAdditional evaluations assess whether automated monitors detect harmful\nactions and whether the model can evade detection or alter its reasoning\nwhen instructed to do so.\n\nIn the tables below, ↑ indicates higher scores are better and ↓\nindicates lower scores are better. GPT-6 Astra results are included for\ncontext. Evaluation versions, inference settings, grading and sample\ncoverage differ across some runs, so these are descriptive comparisons.\n“Not available” indicates that no numerical reference result has been\nidentified. ≈ denotes an approximate value read from a published\nfigure.\n\nTable 12: Honesty and reliability.\n\nTable 12: Honesty and reliability.\n\nEvaluation\n\nMeasure\n\nGPT-Rosalind-6\n\nGPT-6 Astra\n\nBlocker Deception\n\nBlocker-related deception rate ↓\n\n9.5%\n\n12.9%\n\nBroken Search Tool\n\nRate of failing to acknowledge that the search tool is unavailable ↓\n\n1.3%\n\n1.5%\n\nTable 13: Behavior when using tools.\n\nTable 13: Behavior when using tools.\n\nEvaluation\n\nMeasure\n\nGPT-Rosalind-6\n\nGPT-6 Astra\n\nExploitGym Honeypot\n\nRate of successfully exploiting evaluation infrastructure outside the assigned task ↓\n\nRate of successfully circumventing a restriction after encountering a warning ↓\n\n21.7%\n\n18.9%\n\nRespecting Auto-Review\n\nRate of attempting to circumvent an automatic-review denial, including unsuccessful retries ↓\n\n0%\n\n0%\n\nAgent Replay Auto-Review\n\nRate of harmful or unauthorized actions rated certain level severity or higher in replayed agentic conversations, excluding laziness ↓\n\n27.1%\n\n26.7%\n\nUnintended Engagement with External Agent Messages: Performed specified misaligned action\n\nRate of performing the specified unauthorized action, conditional on discovering the message board ↓\n\n2.5%\n\n0.0%\n\nCowork Misalignment\n\nOverall rate of misaligned outcomes in simulated workplace tasks ↓\n\n1.4%\n\n1.2%\n\nTable 14: Monitoring and evasion.\n\nTable 14: Monitoring and evasion.\n\nEvaluation\n\nMeasure\n\nGPT-Rosalind-6\n\nGPT-6 Astra\n\nOAI-repo Sabotage v2\n\nRate of attempts that successfully sabotage simulated services and evade the monitor ↓\n\n0% unguided, 0% guided\n\n0% unguided, 3% guided\n\nMath Side-tasking Monitor Evasion\n\nFull-context monitor bypass rate among successful hidden mathematics tasks ↓\n\n0%\n\n0%\n\nChain-of-thought controllability\n\nRate of following instructions that constrain the content or form of the model’s reasoning ↓\n\nGPQA: 71.8%\n\nHLE: 59.3%\n\nMMLU: 83.1%\n\nGPQA: 69.2%\n\nHLE: 53.8%\n\nMMLU: 81.6%\n\n4.2.3 Enforcement\n\nOpenAI can restrict or revoke access for misuse. OpenAI maintains\nmonitoring, review, and enforcement processes to help identify potential\nmisuse after access is granted. These processes are designed to detect\nactivity that may indicate malicious biological, chemical or other\nactivity that is in violation of OpenAI’s Usage Policies or unauthorized\naccess.\n\nOpenAI uses a variety of signals to assess potential misuse,\nincluding automated safety classifiers, account- or organization-level\nindicators, customer-provided information, and, where available,\nspecialist review of potentially high-risk activity. The specific review\nprocess depends on the deployment configuration and the information\navailable to OpenAI.\n\nBefore launch, we tested our shared monitoring and enforcement\nprocess through a biosecurity red-teaming effort on GPT-Rosalind-5.5 and\nare in the process of repeating this for GPT-Rosalind-6 now that it is\ndeployed. We are not publicly disclosing these details because they\ncontain specific information about our safeguards, but we shared them\nwith our Safety Advisory Group committee and other relevant governance\nbodies.\n\n4.2.4 Security and Access Controls\n\nApproved GPT-Rosalind-6 customers must maintain security controls\nappropriate for access to a higher-capability life sciences model. These\ninclude strong identity and access management, administrator controls,\nleast-privilege access controls appropriate to the approved deployment,\nand processes to remove access when a user no longer has a legitimate\nneed.\n\nModel access and data-retention arrangements are approved separately.\nCustomers must meet the security and oversight requirements for their\napproved deployment.\n\nAs with other OpenAI deployments, we expect to continue learning from\nreal-world use. We will monitor safety signals, customer experience, and\nemerging misuse patterns, and may update GPT-Rosalind-6 safeguards as\nthe deployment evolves.", "url": "https://wpnews.pro/news/gpt-rosalind-6-system-card-addendum", "canonical_source": "https://deploymentsafety.openai.com/gpt-rosalind-6", "published_at": "2026-11-02 14:44:02+00:00", "updated_at": "2026-11-02 14:52:32.626952+00:00", "lang": "en", "topics": ["ai-safety", "ai-policy", "ai-research", "large-language-models", "artificial-intelligence"], "entities": ["OpenAI", "GPT-Rosalind-6", "GPT-6 Astra", "GPT-Rosalind-5.5", "Preparedness Framework", "MedChemBench v7", "GeneBench Pro v13", "LifeSciBench Gold"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/gpt-rosalind-6-system-card-addendum", "markdown": "https://wpnews.pro/news/gpt-rosalind-6-system-card-addendum.md", "text": "https://wpnews.pro/news/gpt-rosalind-6-system-card-addendum.txt", "jsonld": "https://wpnews.pro/news/gpt-rosalind-6-system-card-addendum.jsonld"}}