OpenAI introduces GPT-Rosalind, an AI model designed for research in life sciences.

      Named after the crystallographer who played a crucial role in uncovering the structure of DNA, GPT-Rosalind is OpenAI's inaugural series of models tailored for specific domains, specialized in biochemistry, genomics, and protein engineering. Access is limited to a trusted-access program for approved enterprise clients, including Amgen, Moderna, and Thermo Fisher Scientific.

      On Thursday, OpenAI introduced GPT-Rosalind, a cutting-edge reasoning model created specifically for life sciences research. The model aims to assist with evidence synthesis, hypothesis generation, experimental planning, and multi-step scientific workflows in biochemistry, genomics, and protein engineering, marking OpenAI's first series of models designed for specific purposes.

      It is available as a research preview within ChatGPT, Codex, and the OpenAI API, but access is confined to a trusted-access program for qualified enterprise customers situated in the United States. The model is named in honor of Rosalind Franklin, the British chemist and X-ray crystallographer whose diffraction images of DNA were crucial in revealing the structure of the double helix, and whose contributions were notably absent from the 1962 Nobel Prize awarded to Watson, Crick, and Wilkins.

      This naming serves as a significant acknowledgment: Franklin's work is now widely considered foundational to modern molecular biology, and she is a central figure in discussions concerning the marginalization of women in scientific history.

      OpenAI presents GPT-Rosalind as a means to shorten the duration from scientific concept to clinical evidence. The company estimates that it typically takes about 10 to 15 years to transition a drug from target discovery to regulatory approval in the U.S. GPT-Rosalind is positioned to aid during the initial phases: it can access specialized databases, analyze scientific literature, engage with computational tools, and propose new experimental approaches all within a single interface.

      In conjunction with the model, OpenAI is also launching a Life Sciences research plugin for Codex that links models to over 50 scientific tools and data sources, allowing researchers programmatic access to biological databases and computational workflows.

      Launch partners encompass Amgen, Moderna, Thermo Fisher Scientific, and the Allen Institute. OpenAI is also collaborating with Los Alamos National Laboratory on AI-assisted protein and catalyst design.

      According to benchmark performance data provided by OpenAI, GPT-Rosalind achieved a 0.751 pass rate on BixBench, a bioinformatics benchmark created by Edison Scientific that assesses models on real-world computational biology tasks. On LABBench2, a broader research task benchmark, the model surpassed GPT-5.4 in six out of eleven tasks, with its most notable advantage seen in CloningQA, which involves the complete design of reagents for molecular cloning protocols.

      Remarkably, a third-party evaluation performed with Dyno Therapeutics, a gene therapy company focused on designing AAV capsid proteins, provided striking performance signals. Utilizing unpublished RNA sequences to prevent benchmark contamination, GPT-Rosalind was evaluated on sequence-to-function prediction and sequence generation tasks. The best submission out of ten ranked in the above 95th percentile of human experts for the prediction task and around the 84th percentile for sequence generation, as reported by OpenAI and corroborated by multiple sources covering the launch.

      The rollout comes with significant dual-use warnings that OpenAI has addressed through its access model. Experts have cautioned that AI models trained on biological data might be misused to design harmful pathogens. OpenAI's choice to limit access solely to a vetted trusted-access program—requiring organizations to demonstrate their commitment to enhancing human health and upholding stringent security and governance measures—directly responds to this concern. During the research preview stage, usage will not count against existing API credits.