OpenAI has launched GPT-Rosalind, a specialized AI model dedicated to advancing drug discovery and life sciences research. Named after Rosalind Franklin, the pioneering chemist whose work was crucial in revealing the structure of DNA, this model is the first in OpenAI's series of domain-specific models, finely tuned for fields such as biochemistry, genomics, and protein engineering.
On Thursday, OpenAI announced the debut of GPT-Rosalind, a cutting-edge model crafted specifically for the life sciences sector. This innovative model aims to facilitate various scientific processes, including evidence synthesis, hypothesis generation, experimental planning, and the execution of multi-step scientific workflows.
GPT-Rosalind is available for research preview through platforms like ChatGPT, Codex, and the OpenAI API. However, access is limited to a trusted-access program that caters exclusively to vetted enterprise customers within the United States.
The model's naming serves as a significant recognition of Franklin’s pivotal contributions to molecular biology, particularly her X-ray crystallography work that was essential in uncovering the DNA double helix structure. Franklin's contributions were largely overlooked when the Nobel Prize was awarded in 1962, and the acknowledgment of her name in this model underscores the importance of recognizing women's roles in scientific advancements.
OpenAI positions GPT-Rosalind as a transformative tool designed to shorten the timeline from scientific conception to clinical validation. It is estimated that the current process of bringing a drug from target discovery to regulatory approval in the U.S. takes between 10 to 15 years. By enhancing early-stage research capabilities, GPT-Rosalind is set to streamline various aspects of this lengthy process.
The model boasts the capacity to query specialized databases, interpret scientific literature, interact with computational tools, and propose novel experimental pathways—all integrated into a single user interface. To complement the model, OpenAI has also introduced a Life Sciences research plugin for Codex, which connects researchers to over 50 scientific tools and data sources, granting programmatic access to biological databases and computational pipelines.
Among the launch partners are notable organizations such as Amgen, Moderna, Thermo Fisher Scientific, and the Allen Institute. OpenAI is collaborating with the Los Alamos National Laboratory on projects that leverage AI for protein and catalyst design.
In terms of performance, OpenAI has reported impressive benchmark results for GPT-Rosalind. The model achieved a 0.751 pass rate on BixBench, a bioinformatics benchmark developed by Edison Scientific to assess models on real-world computational biology tasks. Furthermore, on LABBench2, a broader research task benchmark, GPT-Rosalind outperformed GPT-5.4 in six out of eleven tasks, showcasing its most significant advantage in CloningQA, which involves designing reagents for molecular cloning protocols.
A particularly noteworthy performance signal emerged from a third-party evaluation conducted by Dyno Therapeutics, a company specializing in gene therapy and the design of AAV capsid proteins. In this evaluation, GPT-Rosalind was tested on sequence-to-function prediction and sequence generation tasks using unpublished RNA sequences to prevent benchmark contamination. The model's best submissions ranked above the 95th percentile of human experts in the prediction task and around the 84th percentile in sequence generation, as confirmed by multiple outlets covering the launch.
However, the launch of GPT-Rosalind comes with significant dual-use concerns. Experts have cautioned that AI models trained on biological data could potentially be misused to create dangerous pathogens. In response to these risks, OpenAI has chosen to limit access to a vetted trusted-access program. Organizations must demonstrate that they are committed to improving human health outcomes and maintaining robust security and governance controls. During the initial research preview phase, usage of the model will not incur existing API credits.