Ginkgo partners GPT-5 for 36,000 experiments

Ginkgo Bioworks and OpenAI said in February that they had connected GPT-5 to Ginkgo’s cloud laboratory and used it to run more than 36,000 biological experiments with minimal human involvement. The system was used to optimize cell-free protein synthesis, a standard lab process for making proteins outside living cells, according to statements from both companies. Ginkgo said the work reduced costs by 40% relative to the prior benchmark, while OpenAI said the experiments were carried out over six rounds of closed-loop testing. The results were described in a February 5 preprint and in company posts published the same day. ### What exactly did Ginkgo and OpenAI build? OpenAI said GPT-5 was linked to Ginkgo’s cloud lab, an automated wet-lab system that can execute experiments remotely through software. In that setup, the model proposed experiment batches, the lab robots ran them, data came back, and the model used the results to design the next round, according to OpenAI’s description. (prnewswire.com) Ginkgo said the autonomous system designed, executed and learned from experiments with minimal human involvement. The company’s January 13 news release said the work covered 36,000 experimental conditions across six iterative cycles in Boston. (openai.com) ### What process were they trying to improve? The target was cell-free protein synthesis, or CFPS, which researchers use to make proteins in a reaction mix rather than inside cells. OpenAI said the project focused on lowering the cost of producing superfolder green fluorescent protein, a commonly used benchmark protein. (prnewswire.com) The bioRxiv preprint said the autonomous lab optimized the cost efficiency of CFPS and reported a 40% reduction in specific cost relative to the state of the art. The paper also said the system pursued both cost and titer, or output, during iterative optimization. ### Where does the 40% figure come from? OpenAI said the system reduced the cost of protein production from $698 per gram to $422 per gram in the benchmark task. (openai.com) RD World, citing the company materials, reported the same figures and said the work unfolded over about six months. Ginkgo’s release described the result as a 40% improvement over the state-of-the-art scientific benchmark. (biorxiv.org) The preprint uses similar language, saying the reduction was achieved relative to the prior state of the art rather than against a company-defined internal baseline. (openai.com) ### What does “10-100x faster” refer to? The “10-100x” timeline claim does not appear in the Ginkgo press release, the OpenAI post surfaced here, or the bioRxiv abstract that describes the experiment. Those sources verify the 36,000 experiments, the six closed-loop rounds and the 40% cost reduction, but not that specific speed range. (prnewswire.com) An X thread on May 22 cited the Ginkgo-OpenAI work as part of a broader bio-and-AI optimism post. But the underlying primary materials available from Ginkgo, OpenAI and bioRxiv support the cost and experiment-count claims more clearly than the “10-100x” wording. (prnewswire.com) ### Why did this show up again in May? An X thread circulated on May 22 that grouped the Ginkgo-OpenAI project with other recent biotech and AI advances. That appears to be why the partnership resurfaced in social discussion months after the original February disclosures. (prnewswire.com) The next public source for details remains the February 5 preprint, along with the company write-ups from Ginkgo and OpenAI. Those documents name the participants, describe the six experimental rounds and provide the benchmark cost figures used in the announcement. (openai.com) (prnewswire.com)