OpenAI Presented GPT-Rosalind — an AI Model for Biology, Genomics, and Drug Development

On April 16, OpenAI introduced GPT-Rosalind—the company's first specialized model for life sciences, covering tasks at the intersection of biology, pharmaceuticals, and applied medicine. The model is named in honor of Rosalind Franklin and is conceived not as a replacement for scientists, but as a tool to accelerate the most demanding aspect of scientific work: literature review, data verification, hypothesis formation, and planning next steps. For context, the path from identifying a new drug target to FDA approval typically takes 10 to 15 years.

OpenAI is betting that AI can shorten at least the early stages of this cycle, where researchers spend enormous amounts of time on analysis, fact-checking, and working with fragmented sources. According to the company's description, GPT-Rosalind is optimized for scientific processes requiring reasoning about chemistry, proteins, genes, biological pathways, and experimental protocols. Unlike universal chatbots that can do a little of everything, the focus here is on long research chains: gathering and comparing dozens of papers, retrieving data from specialized databases, interpreting results, proposing a working hypothesis, and helping plan the next experiment.

OpenAI emphasizes separately that this is not simply about text generation. The model should help summarize the evidence base, generate hypotheses, plan experiments, and analyze data—in other words, operate within real research workflow, not just answer questions in a chat window. Alongside this, they released the Life Sciences research plugin for Codex.

The plugin provides access to more than 50 scientific tools, literature sources, multi-omics databases, and biological services. In practice, this means a researcher can run sequence searches, view protein structures, gather publications, find open datasets, and link all of this to the model's reasoning—all in one interface. For laboratories and pharmaceutical companies, this matters more than just another LLM: value emerges where the model not only formulates an answer, but can access the necessary data and integrate into existing computational workflows.

According to OpenAI's public benchmarks, GPT-Rosalind shows the best result on BixBench among models with published metrics. On LABBench2, the model outperformed GPT-5.4 in six of eleven tasks; the company saw the most significant improvement in CloningQA, where designing reagents and a complete plan for molecular cloning is required.

OpenAI separately tested the model with Dyno Therapeutics on the task of predicting and generating RNA sequences by function, using unpublished data that could not have entered the training set. In the Codex environment, the model's best results out of ten attempts exceeded the 95th percentile of human experts on the prediction task and were roughly at the 84th percentile on the sequence generation task. This does not prove that AI is ready to make discoveries on its own, but shows that in narrow bioinformatics scenarios, it is beginning to compete with strong specialists.

The launch has been cautious. GPT-Rosalind is currently available as a research preview through ChatGPT, Codex, and API, but only for qualified corporate clients in the United States under a trusted access program. OpenAI cites enhanced security measures, access control, and governance and internal control requirements: an organization must conduct legitimate research with clear benefits to human health, restrict access to the model, and follow rules to prevent misuse.

Among partners and early users, the company names Amgen, Moderna, Thermo Fisher Scientific, Allen Institute, and other organizations from the life sciences ecosystem. In parallel, OpenAI is working with Los Alamos National Laboratory on AI support for protein and catalyst design. For the market, this is an important shift.

OpenAI demonstrates that the next big bet is not only more general and larger models, but narrowly specialized systems tailored to specific professional processes. In life sciences, the cost of error is high, data is complex, and success is measured not by the number of polished answers, but by the quality of hypotheses and speed of experiments. If GPT-Rosalind truly takes root in laboratory processes, it will reinforce the trend toward vertical AI models for industries where value is born from deep domain understanding, access to tools, and work in a tightly controlled environment.