GPT-Rosalind

GPT-Rosalind is a specialized reasoning model developed by OpenAI for applications in the life sciences, drug discovery, and translational medicine. It is the first model in a planned Life Sciences series and was introduced on April 17, 2026, as a research preview for select enterprise customers in the United States under the Trusted Access program. The model's name references British chemist and crystallographer Rosalind Franklin, whose X-ray diffraction research contributed to the discovery of the double helix structure of DNA.

The model was built on OpenAI's latest internal models and fine-tuned specifically for deep biological analysis. OpenAI trained it on 50 of the most common biological workflows and integrated access to major public biological databases. GPT-Rosalind supports multi-step scientific workflows, including literature evidence synthesis, biological hypothesis generation, experiment planning, and genomic and biochemical data analysis. It can query specialized databases, process scientific literature, interface with computational tools, and suggest new experimental pathways within a single interface.

The model is available in ChatGPT, Codex, and via the OpenAI API exclusively to qualified enterprise customers in the US who have passed a security review and demonstrated that they conduct legitimate research with clear public benefit. During the research preview phase, usage does not draw from existing credits or tokens. Alongside the model, OpenAI released a free Life Sciences Research Plugin for Codex, available on GitHub, connecting models to over 50 scientific tools and databases — including AlphaFold, Bgee, and BindingDB — covering human genetics, functional genomics, protein structure, biochemistry, and clinical research discoveries.

Partners and applications

Confirmed partners and customers collaborating on the model at launch include Amgen, Moderna, Novo Nordisk, Thermo Fisher Scientific, Oracle Health and Life Sciences, NVIDIA, Allen Institute, Benchling, and UCSF School of Pharmacy. OpenAI is also collaborating with Los Alamos National Laboratory on AI-assisted protein and catalyst design, and Dyno Therapeutics participated in the model evaluation phase. Consulting firms McKinsey & Company, Boston Consulting Group, and Bain & Company are associated with OpenAI's Life Sciences team.

Benchmark results

On the BixBench benchmark — which evaluates bioinformatics tasks and data analysis in real research environments and was developed by Edison Scientific — GPT-Rosalind achieved a Pass@1 score of 0.751, the highest among models with published results. For comparison: GPT-5.4 scored 0.732, Grok 4.2 scored 0.728, GPT-5.2 scored 0.698, GPT-5 scored 0.611, and Gemini 3.1 Pro scored 0.550 (OpenAI data). On the LABBench2 benchmark — covering literature search, database access, sequence manipulation, and protocol design — GPT-Rosalind outperformed GPT-5.4 on 6 of 11 tasks, with the largest margin in the CloningQA category, which requires comprehensive design of DNA and enzymatic reagents for molecular cloning protocols. In OpenAI's internal tests, the model outperformed GPT-5, GPT-5.2, and GPT-5.4 across all five categories: chemistry, biochemistry and protein understanding, phylogenetics, experiment design and analysis, and tool use.

In an evaluation conducted in collaboration with Dyno Therapeutics on unpublished, never-before-seen RNA sequences — to exclude training data memorization — the model's best-of-ten submissions ranked above the 95th percentile of human experts on a sequence-to-function prediction task and at the 84th percentile on sequence generation. This evaluation was benchmarked against a set of 57 historical results from AI and biology experts.

Access and safety

Given the model's dual-use potential regarding biological threats, OpenAI has imposed strict access restrictions. Organizations applying for access must pass a security review and demonstrate three conditions: conducting legitimate scientific research with clear public benefit, having appropriate governance, compliance, and misuse-prevention procedures in place, and restricting access exclusively to approved users in secure, managed environments. OpenAI has built systems into the model for detecting potentially dangerous activities. Pricing and broad availability will be announced after the research preview phase concludes.