OpenAI launches GPT-Rosalind for drug discovery; US access limited

OpenAI unveiled GPT-Rosalind on Thursday, a reasoning model for life sciences designed to speed early-stage drug discovery. The company said access will be limited to vetted U.S. enterprise customers that pass a qualification and safety review.

OpenAI described the model as a tool to help researchers process scientific literature, query databases, design reagents and interpret experimental results more quickly than standard workflows allow. The company framed the model as aimed at shortening time spent on common, time-consuming tasks in early drug development.

OpenAI released benchmark results showing gains over its prior general models on several biology tasks. On BixBench, a suite of real-world bioinformatics problems, GPT-Rosalind logged a 0.751 pass rate, the highest among models with published results. In LABBench2 evaluations the model outperformed OpenAI’s GPT-5.4 on six of 11 tasks and outscored GPT-5.4 in every life-science case the company assessed. On sequence work, the model’s best-of-ten predictions ranked above the 95th percentile of human experts; on sequence generation its performance was around the 84th percentile.

OpenAI said it will work with Dyno Therapeutics to test the model against unpublished RNA sequences to check for memorization. The company also noted that during the research preview use of GPT-Rosalind will not consume existing API credits.

Alongside the model, OpenAI is releasing a free Life Sciences research plugin for its Codex environment that connects to more than 50 scientific databases and tools, including protein structure lookup, sequence search and genomics pipelines. Enterprise customers approved for GPT-Rosalind will receive the model’s specialized reasoning layer; other users can access the plugin with standard models.

OpenAI named several industry partners for the launch, including Amgen, Moderna and Thermo Fisher Scientific, and said it is collaborating with Los Alamos National Laboratory on AI-guided protein and catalyst design. The company previously released the Prism scientific writing workspace in January as an initial research workflow product.

Access to GPT-Rosalind is restricted to U.S. enterprise customers who complete a qualification and safety review. OpenAI cited concerns about the potential misuse of biological data. An international group of more than 100 scientists has urged tighter controls on biological datasets used to train AI models because of risks tied to pathogen design and other dual-use capabilities.

Joy Jiao, OpenAI’s life sciences research lead, described the model as not capable of creating new treatments on its own and said the company expects it to speed parts of the research process: “We do think there’s a real opportunity to help researchers move faster through some of the most complex and time‑intensive parts of the scientific process.”

The model is named for British chemist Rosalind Franklin, whose X-ray crystallography work contributed to the discovery of DNA’s double helix. No drug discovered entirely by AI has yet cleared phase‑3 clinical trials.