OpenAI unveils AI chemist GPT-Rosalind and a 750-task test
On the same day, OpenAI released two things from the life sciences: a near-autonomous AI chemist called GPT-Rosalind, and a new benchmark, LifeSciBench, that checks how much models can really do real scientific work.
- On the same day, OpenAI released an AI chemist, GPT-Rosalind, and a new benchmark, LifeSciBench.
- The benchmark holds 750 expert-written tasks across seven workflows and seven biology domains.
- It was built by 173 scientists with industry experience.
- In testing, GPT-Rosalind improved a hard medicinal-chemistry reaction.
Can artificial intelligence actually do real lab work, not just talk about it? On June 17, OpenAI tried to answer that question on two fronts at once. It released GPT-Rosalind, a model meant to act as a near-autonomous chemist, and LifeSciBench, a benchmark that checks how much models can really do scientific work from start to finish.
The distinction matters. Asking a model to write an essay about a reaction is one thing. Asking it to genuinely improve that reaction, step by step, the way a person in a lab would, is another. OpenAI says that in testing, GPT-Rosalind improved a difficult reaction in medicinal chemistry, the field that builds the molecules that later become drugs.
What exactly does LifeSciBench measure?
LifeSciBench may tell us more, as onlookers, than the model itself. It is a benchmark, an exam by which models are graded. It contains 750 tasks written by experts, spread across seven workflows and seven biology domains. It was built by 173 scientists with industry experience, which means the tasks are not lifted from a textbook but resemble what real research actually involves.
The benchmark grades six categories of scientific work. The point is to see where a model gets stuck, not just how many correct answers it gives to isolated questions. Real lab work is not a quiz. It is a chain of decisions that build on one another, where one early mistake can topple the whole sequence.
GPT-Rosalind also has its own narrower test, called MedChemBench. It sticks to medicinal chemistry and checks a few concrete things: whether the model understands chemical structures, whether it can predict a molecule’s potency, its toxicity, and its so-called ADME properties, meaning how a drug behaves in the body, from absorption to elimination. On that benchmark, GPT-Rosalind posts leading results. Our advice is to read those numbers as vendor figures until someone independent reproduces them.
Why now, and what does it change for us?
The trend has been clear for a while. After models that write text and code, the next wave is models that are supposed to help with actual research. Life sciences are an attractive target because the money is large and the patience required is enormous, since developing a single drug takes years. If AI shortens even part of that path, the effect is measurable.
But there is also a reason for restraint. A benchmark built by the same company that sells the model is not an independent verdict. It helps to have a shared exam like LifeSciBench, because other models can then be measured on the same task. Only when independent groups reproduce the results, and when chemists in real labs say the model genuinely saves them time, will we know what this is worth beyond a press release.
For now, we have a clear picture of where the industry is heading. AI is describing the world less and trying to do something in it more. So would you trust a drug whose key reaction was partly designed by a model, as long as it passed all the same checks as any other?
