OpenAI Upgrades GPT-Rosalind, Its Drug-Discovery Model

OpenAI has given its life-sciences model a substantial upgrade, and the headline figure is an unusual one for a frontier lab to lead with. The new GPT-Rosalind does more while using 31 percent fewer tokens than GPT-5.5, the general-purpose model it is built on.

GPT-Rosalind, launched earlier this spring and named for Rosalind Franklin, whose X-ray work was central to understanding the structure of DNA, is OpenAI's attempt to build a model that earns its keep in a laboratory rather than a chat window. The 3 June update folds in GPT-5.5's agentic coding and tool use, then layers stronger performance in the domains that matter to drug hunters: medicinal chemistry, genomics and quantitative biology.

What it actually does

In OpenAI's own evaluations, the model posts broad gains on tasks written by working biologists, from complex medicinal-chemistry queries to wet-lab troubleshooting, the unglamorous business of working out why an experiment misbehaved. The token efficiency matters more than it sounds. Research workflows can call a model thousands of times over a single project, so a third off the token bill is the difference between an experiment a lab can afford to run and one it cannot.

A note of caution is warranted on the benchmarks. These are OpenAI's figures, measured on OpenAI's chosen tasks, and they have not yet been reproduced independently. The model is rolling out as a research preview to eligible organisations rather than as an open product, which is the right posture for a tool that touches genomics and pathogen biology.

The biodefense question

That last point is why OpenAI paired the release with Rosalind Biodefense, an initiative aimed at defensive applications in the life sciences. A model fluent in medicinal chemistry and viral genomics is dual-use by nature. The same reasoning that helps design a vaccine can, in the wrong hands, help design something far worse. Keeping the rollout controlled, and standing up a biodefense programme alongside the model, is OpenAI's answer to a risk the whole field is now wrestling with.

The broader trend here is specialisation. The first era of large models rewarded generality, one system that could do a little of everything. The interesting work in 2026 is increasingly about narrow, expert models tuned for a single demanding field and built to run cheaply enough that a research group can keep one going all day. A vaccine designed by AI reached its first human trial only last week. Tools like GPT-Rosalind are how that pipeline gets faster.