OpenAI 发布 GPT-Rosalind-5.5 系统卡，定位生命科学高能力模型并设可信访问门槛

2. Model Data and Training

GPT-Rosalind-5.5 is incrementally trained from GPT-5.5 with training data related to beneficial life sciences capability. Unlike GPT-5.5, it is trained not to refuse sophisticated biology queries, and leverages a trusted access and responsible deployment structure as the primary safeguard.

3.2 Genomics and quantitative biology

To evaluate GPT-Rosalind-5.5 on genomics understanding and quantitative biology, we designed an agentic evaluation on long horizon, end-to-end genomics analysis tasks. This evaluation assesses agentic performance on long-horizon quantitative tasks: based on realistic scientific data, can an agent plan valid analysis, QC, modeling, and corrections to arrive at decision-relative answers? Genebench spans a variety of domains, including functional genomics, spatial transcriptomics, proteomics, epigenomics, and applied genetics.

4.1.1 Biological and Chemical

As we did for GPT-5.5, we are treating this launch as High capability in the Biological and Chemical domain. Below, we describe the evaluations we used to reach this conclusion. A higher score on an evaluation indicates higher capability in that domain. As expected, GPT-Rosalind-5.5 scores meet or exceed GPT-5.5’s capability of all the following evaluations but one (Multi-Select Virology Troubleshooting). We also show values for GPT-5.5 Pro, which is run with roughly an order of magnitude higher inference budget and thus is expected to exceed GPT-Rosalind-5.5 performance in some areas. In particular, GPT-Rosalind-5.5 was not specifically trained on virology or biothreat creation, and for these evaluations it expectedly underperforms GPT-5.5 Pro.

4.2.2 Model-level boundaries against harmful biological and cyber assistance

4.2.2.1 Biological and Chemical boundaries

GPT-Rosalind-5.5 is designed to be more useful for legitimate advanced biological research than generally available ChatGPT models. This includes support for advanced life sciences workflows such as literature and knowledge retrieval, data parsing and analysis, computational biology, and use of specialized biology tools and databases.

Like GPT-5.5, GPT-Rosalind-5.5 is trained to refuse malicious requests that would meaningfully enable biological weaponization.

Because some advanced biological capabilities are dual-use, GPT-Rosalind-5.5 is not broadly available. Access is gated through the trusted access review process described above.

4.2.3 Monitoring, Review, and Revocation

OpenAI maintains monitoring, review, and enforcement processes to help identify potential misuse after access is granted. These processes are designed to detect activity that may indicate biological weaponization, malicious cyber activity, unauthorized access, or use that is materially inconsistent with the customer’s approved purpose.

OpenAI uses a variety of signals to assess potential misuse, including automated safety classifiers, account- or organization-level indicators, customer-provided information, and, where available, specialist review of potentially high-risk activity. The specific review process depends on the deployment configuration and the information available to OpenAI.

Because advanced life sciences work is inherently dual-use, some legitimate research activity may require contextual review. Our process considers both the risk of the activity and the apparent purpose of the use, including whether it remains aligned with the customer’s approved life sciences, public-health, biosafety, biosecurity, or biodefense use case. Activity that appears to cross into biological weaponization, malicious cyber activity, unauthorized access, or other prohibited harmful use may be flagged for review and action. OpenAI also seeks to monitor for unresolved customer security concerns, failure to maintain required access controls, or material changes in approved use case, deployment pathway, or user population.

OpenAI reserves the option to apply additional monitoring, request information from the customer, narrow or restrict access, require remediation of security or governance issues, or revoke GPT-Rosalind-5.5 access.

4.2.4 Security and Access Controls

Approved GPT-Rosalind-5.5 customers must maintain security controls appropriate for access to a higher-capability life sciences model. These include strong identity and access management, administrator controls, role-based access controls, and processes to remove access when a user no longer has a legitimate need.

These controls are intended to help ensure that GPT-Rosalind-5.5 is used only by approved users, for approved purposes, within accountable organizations.

As with other OpenAI deployments, we expect to continue learning from real-world use. We will monitor safety signals, customer experience, and emerging misuse patterns, and may update GPT-Rosalind-5.5 safeguards as the deployment evolves.