GPT-Rosalind: OpenAI Launches Its First AI Model Dedicated to Biology and Drug Discovery
OpenAI has launched GPT-Rosalind, its first frontier model specifically designed for research in the field of biological sciences. The launch, which took place on Thursday, April 16, introduces a LLM trained on 50 common biological workflows and integrated with major public databases in the sector, with the stated goal of accelerating early discovery phases in pharmacological and genomic research.
GPT-Rosalind: Why a Model Dedicated to Biology
With this model, OpenAI aims to address a structural problem in biological research: researchers are overwhelmed by volumes of data that no single expert can master: scientific literature, sequencing outputs, gene expression data, protein profiles, experimental results. Added to this is the extreme fragmentation of sub-disciplines, where the technical vocabulary of one research group can seem almost opaque to another. Yunyun Wang, Life Sciences Product Lead at OpenAI, framed the launch in exactly these terms: GPT-Rosalind was born to bridge the gap between the available data and human capacity to synthesize it into usable hypotheses.
The traditional drug discovery process typically takes 10-15 years from target discovery to regulatory approval. A significant portion of this time is wasted in preliminary phases: linking papers, molecules, proteins, genes, and lab results into a coherent hypothesis. GPT-Rosalind is designed to operate exactly in this space, integrating reasoning over literature, databases, and experimental data into multi-step workflows.
Architecture and Capabilities of the New GPT-Rosalind
OpenAI built GPT-Rosalind starting from a general LLM backbone, then applying additional training and tuning aimed at the experimental and analytical workflows of biology. The model integrates procedural knowledge and scientific literature in various disciplines such as sequence analysis, expression profiling, and protein biochemistry. OpenAI's new model is capable of connecting genotype to phenotype through known pathways and regulatory mechanisms, inferring structural and functional properties of proteins, and suggesting follow-up experiments within a single workflow.
One technical element that OpenAI explicitly highlighted is the so-called skepticism tuning: the model has been specifically calibrated to reduce sycophancy and overconfidence, two of the most critical issues when applying LLMs to high-risk scientific domains. In other words, when presented with a target of little value, the model is designed to directly reject the proposal instead of indulging the user. This is a design choice that reflects an awareness that hallucinations in the context of biological sciences are not just an accuracy issue, but can have material implications on research decisions.
In terms of performance, OpenAI reports that GPT-Rosalind ranks highly on BixBench benchmarks and surpasses GPT-5.4 on 6 of the 11 tasks of LABBench2, with the widest margin on the cloning design workflows, which is a practical test of the ability to select the correct DNA parts and enzymes for a lab protocol. In an evaluation conducted with Dyno Therapeutics, the model's best submissions in the Codex app placed above the 95th percentile of human experts on an RNA prediction task, and around the 84th percentile on a gene sequence generation task.
GPT-Rosalind: Access and Distribution
The model is available as a research preview in closed access, initially reserved for qualified Enterprise clients. Delivery occurs through ChatGPT Enterprise, Codex, and the API, with enterprise governance controls, Regulated Workspaces, BAA, SOC 2 Type 2 compliance, and HIPAA-aligned standards. OpenAI has emphasized that it does not train its models on customer data. Among the first access partners are the pharmaceutical giant Amgen, vaccine manufacturer Moderna, and the Allen Institute, the non-profit bioscience research organization founded by Paul Allen.
Alongside the full model in restricted access, OpenAI has made a Life Sciences Research Plugin available in general access, designed as an entry point for teams that do not yet fall under the enterprise program. The competitive positioning is direct: unlike Google's and Microsoft's approach, which focus on generic scientific models, OpenAI has chosen to verticalize on a specific domain, directly addressing the operational difficulties of research laboratories. The chosen name, Rosalind, is a tribute to Rosalind Franklin, the British chemist whose work was pivotal to the discovery of the DNA structure and also signals the strategic direction that OpenAI intends to take in the scientific field.
Limitations and Open Questions
Closed access is both a control element and a limitation: without publicly verifiable benchmarks, shared validation datasets, and access to the broader scientific community, it is difficult to independently evaluate OpenAI's claims regarding performance. The model does not release weights nor publishes detailed benchmark tables. For teams planning to integrate it into their workflows, critical questions remain regarding the transparency of training sources, negative controls, confidence calibration, and integration with existing electronic lab notebooks and LIMS systems, that is, the software solutions for managing laboratory tasks.