Ginkgo Bioworks' Autonomous Laboratory Driven by OpenAI's GPT-5 Achieves 40% Improvement Over State-of-the-Art Scientific Benchmark

Ginkgo Bioworks has announced an impressive milestone in collaboration with OpenAI, showcasing an AI-driven system that autonomously conducts and learns from biological experiments with minimal human input. The new preprint reveals that this system has successfully reduced the costs of cell-free protein synthesis reactions by 40% compared to existing methods, while executing an astounding 36,000 experimental conditions over six iterative cycles.

In this pioneering study, Ginkgo merged OpenAI's GPT-5 reasoning model with its own cloud laboratory infrastructure, which incorporates reconfigurable automation carts (RAC) and Catalyst automation software. This combination allowed the AI to design, perform, and analyze experiments in an iterative, closed-loop fashion. GPT-5 utilized internet access, a computer equipped with data analysis software, and previous experimental data to operate as an experimental scientist. Over six months, it designed cost-effective cell-free protein synthesis compositions that surpassed previous scientific literature.

"By pairing a frontier large language model with an autonomous lab, we found reaction compositions that are notably cheaper than prior state of the art," stated Reshma Shetty, co-founder of Ginkgo Bioworks and co-author of the study. "We expect more and more experiments to be run on autonomous labs where reagent and consumables costs dominate the cost of an experiment. Lower cost reagents for protein production enable more data generation and thus more scientific progress per dollar spent."

The autonomous lab produced a standard benchmark protein, superfolder green fluorescent protein (sfGFP), at a total reaction component cost of $422 per gram, significantly lower than the previously recorded $698 per gram, marking a 40% reduction. The high material costs and complex optimization associated with cell-free protein synthesis made it an ideal candidate for this autonomous experimentation approach.

"At OpenAI, this was the first time we were able to interface a frontier model with an autonomous lab to carry out experimentation at a very large scale," remarked Joy Jiao, life sciences research lead at OpenAI and co-corresponding author of the study. "This success points to how AI systems can augment the experimental workflow, contributing to hypothesis generation, testing, and refinement based on real-world data."

The autonomous lab completed over 580 384-well plates, tested 36,000 reaction compositions, and generated nearly 150,000 experimental data points. Human roles were primarily limited to preparing reagents, overseeing the system, and managing loading and unloading tasks, while the GPT-5-driven lab handled experimental design, execution, data interpretation, and hypothesis generation. Remarkably, the AI also identified and prioritized new reagents for testing, some of which aligned with findings from published research it hadn't been trained on.

To ensure the AI did not propose impractical or invalid experiments, each design underwent validation against a Pydantic model prior to execution, checking for plate layout, standards, controls, replication, reagent availability, and volume constraints. Only validated experiments could proceed, with additional scoring emphasizing scientific rigor and consideration of past results. GPT-5 also generated comprehensible lab notebook entries documenting its analyses, observations, and rationale, ensuring transparency in its decision-making process.

"This is AI doing real experimental science: designing experiments, running them, and learning from the results," said Jason Kelly, co-founder and CEO of Ginkgo Bioworks. "AI combined with autonomous labs is needed to keep the United States competitive in science worldwide – the recently announced Genesis Mission led by the U.S. Department of Energy to bring AI into science is leading the way here and I'm excited that our results with OpenAI show this approach is working."

The Pydantic model will be released as open source, and the AI-optimized cell-free reaction mix can be ordered by the scientific community via Ginkgo's reagent store. The findings are also detailed in a preprint that is yet to undergo peer review, with the full manuscript available on OpenAI's website and forthcoming on bioRxiv.