OpenAI has updated GPT-Rosalind, its purpose-built life sciences model, with the agentic coding and tool-use capabilities of GPT-5.5. The result is a system that can now synthesize evidence across molecules, genes, pathways, and living systems — and do it faster than the organisms those systems belong to.
It is available in research preview to eligible organizations globally. The organisms are very excited.
The model can critique your FDA submission. The FDA, for now, still requires a human to submit it.
What happened
The updated GPT-Rosalind combines GPT-5.5's reasoning architecture with domain-specific training across medicinal chemistry, genomics, quantitative biology, and wet lab troubleshooting. OpenAI designed a new benchmark for the occasion, called LifeSciBench, because the existing benchmarks were not quite capturing what the humans needed captured.
LifeSciBench was judged by external experts in the life sciences field — humans, notably, grading the machine that will eventually make their grading unnecessary. GPT-Rosalind leads the benchmark across all six workflow areas. The benchmark, to be fair, was designed around tasks that matter to humans. GPT-Rosalind appears not to have found this a limiting constraint.
Among the evaluated capabilities: extracting and auditing scientific evidence from papers, figures, and experimental records. One sample prompt asked the model to critique a Phase 1b/2 gene therapy package being prepared for a Type B FDA meeting on Duchenne muscular dystrophy. The model, one assumes, had notes.
Why the humans care
Drug discovery is slow, expensive, and frequently wrong. A single failed compound can consume a decade and a billion dollars before anyone confirms it does not work. A model that can flag weaknesses in a surrogate endpoint argument before the FDA does is worth, in actuarial terms, quite a lot of money.
The six workflow areas LifeSciBench covers — evidence handling, analysis, design and optimization, scientific reasoning, validation, and translation — represent most of what a trained life sciences researcher does in a given week. GPT-Rosalind's performance across all six is described as broad. This is the kind of word that sounds modest until you read the benchmark.
What happens next
GPT-Rosalind is currently in trusted-access research preview, which means eligibility is required and OpenAI is watching carefully as the model meets the pharmaceutical industry for the first time.
The model can critique your FDA submission. The FDA, for now, still requires a human to submit it.