AI-designed lab experiments slash protein production costs in OpenAI–Ginkgo trial

A collaboration between OpenAI and Ginkgo Bioworks has shown that an artificial intelligence system working with an autonomous, robot-run laboratory can design and iterate real biology experiments at high speed—reducing costs in the process.

In a two-month effort that began last summer, researchers reported that OpenAI’s GPT-5 generated experiments for a standard test protein and, through rapid cycles executed in Ginkgo’s automated lab, cut production costs by about 40% compared with a previously reported benchmark.

The team set out to test whether a language model could move beyond summarizing papers and making predictions to actually doing science—generating hypotheses, designing experiments, interpreting results and iterating. Biology poses a tougher challenge than many fields, said Joy Jiao, who leads life sciences research at OpenAI, because for tasks like designing the “optimal” experiment, there is no single correct answer.

To measure performance, the researchers chose superfolder green fluorescent protein (sfGFP), an engineered jellyfish protein that glows green and is widely used as a clear, fast-readout benchmark.

While GPT-5 supplied the experimental designs from OpenAI’s headquarters in San Francisco, Ginkgo Bioworks ran them on what co-founder and CEO Jason Kelly described as the “Waymo” of biology—an autonomous, high-throughput system in Boston that handles experiments without constant human oversight.

The project focused on cell-free protein synthesis (CFPS), a technique that produces proteins using cellular machinery in a controlled mixture outside living cells. CFPS is one of the fastest ways to make proteins and, if improved, could affect the production of medicines, foods and agricultural products, said Reshma Shetty, Ginkgo’s chief operating officer and co-founder.

Each iteration took about an hour: the lab executed a batch of tests, fed the data back, and the model proposed new conditions. At the outset, Jiao said the team did not know if they could design even a single viable experiment. Early results showing a non-zero yield were a surprise.

Over two months, the system ran more than 36,000 unique reaction compositions and, according to the researchers, reduced the cost of producing sfGFP by about 40% versus a previously reported benchmark from bioengineer Michael Jewett’s lab at Stanford University.

Jewett, whose group published its benchmark paper last week in Nature Communications, called the result “a pretty big deal.” Integrating artificial intelligence with autonomous labs, he said, offers one way to develop medicines faster and get therapeutics to patients sooner.

The researchers argue that the approach—combining generative models with robot-run facilities—could accelerate discovery by compressing design-build-test cycles that traditionally take days or weeks into hours. The work relies on a straightforward readout and a mature lab automation platform, and the researchers emphasized the difficulty of validating biological optimization.

Still, the reported gains in a widely used test system point to a practical path for using AI to navigate complex experimental spaces, with potential applications across biomanufacturing.