Google is rolling out Gemini for Science, a set of experimental tools designed to compress scientific work that would typically take months or years into a matter of days. The program spans Google Research, DeepMind, Google Cloud, and Google Labs, and puts three prototypes into the hands of working researchers through Google Labs. It also extends to enterprise customers through Google Cloud, with early partners including BASF, Klarna, Bayer Crop Science, and the US National Labs as part of the Department of Energy’s Genesis Mission.
A central piece is Co-Scientist, detailed in a new paper in Nature. It is a team of Artificial Intelligence agents built on Gemini that work together like a research group, with separate agents generating ideas, critiquing them, ranking them, improving them, and reviewing the overall process, while a lead agent manages the workflow. Google Research positions this as a case for general-purpose agents rather than narrow specialized models. Yossi Matias said Artificial Intelligence acts as an amplifier of human ingenuity and described a partnership with Imperial College in which researchers had spent years arriving at a bacterial hypothesis that Co-Scientist reached in days.
Google says the system is already being applied in several research settings. At Stanford, researchers used Co-Scientist to help find an existing drug that reduced signs of liver scarring in lab-grown tissue. The platform is also being used at Calico Life Sciences to study aging, at the University of Edinburgh to look for new liver disease treatments, and at the University of Cambridge to study how viruses like the flu and COVID-19 can spread from animals to humans.
The broader Gemini for Science package includes Hypothesis Generation through Co-Scientist, Computational Discovery built on AlphaEvolve and ERA, Literature Insights powered by NotebookLM, and Science Skills that connect agentic tools to more than 30 life science databases including UniProt, AlphaFold Database, and AlphaGenome API. Lizzie Dorfman said that in one epidemiological forecasting project, her team generated 200,000 candidate models. Most were discarded quickly, but the ability to explore that volume changed the workflow and increased what a single researcher could accomplish.
