It had been thought that AI would play a big part in digital-only domains like math and coding, but cleverly connecting these systems to the physical world is also producing some interesting results.
In a collaboration with Ginkgo Bioworks, OpenAI has demonstrated that its frontier model GPT-5 can autonomously design and execute laboratory experiments at scale, achieving a 40% reduction in the cost of producing proteins through cell-free protein synthesis. The project represents a significant step toward AI systems that can not only reason about scientific problems but also drive physical experimentation in real-world lab environments.
The system worked by creating a closed loop between GPT-5 and Ginkgo’s cloud laboratory—a fully automated wet lab controlled through software where robots handle the physical work of running experiments. GPT-5 would propose batches of experiments, the robotic lab would execute them, and the results would feed back to the model to inform the next round of testing. Over six iterations spanning two months, this setup tested more than 36,000 unique reaction compositions across 580 automated plates.
Cell-free protein synthesis is a technique for producing proteins without growing living cells, instead running the protein-making machinery in a controlled mixture. It’s widely used for rapid prototyping in drug development, diagnostics, and industrial applications, but optimizing it is notoriously difficult. The process requires balancing dozens of interacting components—DNA templates, cell lysates, energy sources, salts, and other biochemical ingredients—and standard formulations tend to be expensive at the scale autonomous labs can achieve.
What made the approach effective was the sheer volume of experimentation possible when AI drives the process. While a human team might run dozens of reactions in the time the autonomous system completed thousands, this throughput allowed GPT-5 to identify patterns and optimal combinations that would be easy to miss in manual workflows. The model found reagent mixtures that hadn’t been previously tested in this configuration and discovered compositions that performed better under the specific constraints of high-throughput, plate-based automation—such as lower oxygen availability compared to traditional test tube experiments.
The improvements came from exploring parameters that aren’t always prioritized in manual optimization, including buffering systems, energy regeneration components, and polyamines. At high throughput, these became testable variables rather than background assumptions. The system also identified a key insight about the cost structure itself: with lysate and DNA now dominating expenses in cell-free protein synthesis, maximizing protein yield per unit of expensive input became the highest-leverage strategy for reducing overall costs.
OpenAI implemented strict validation to ensure the loop remained grounded in physical reality. Before any experiment could run, programmatic checks verified that AI-designed protocols were actually executable on the automation platform, preventing what the team calls “paper experiments”—procedures that might look reasonable in text but can’t be carried out by robots.
The work builds on OpenAI’s earlier demonstrations of GPT-5 improving wet-lab protocols through closed-loop experimentation, but this project specifically targeted cost reduction in a widely used industrial process. The 40% reduction in production costs included a 57% improvement in reagent costs and yielded novel reaction compositions more robust to the conditions common in autonomous laboratory settings.
The implications extend beyond protein synthesis. As proteins underpin many modern medicines, diagnostics, research tools, and industrial enzymes, making their production faster and cheaper could accelerate the path from early research to practical applications. More broadly, the project suggests that the bottleneck in life sciences—where progress depends on physical experimentation that takes time and money—may start to ease as AI systems gain the ability to autonomously propose, execute, and learn from experiments at scales impossible for human teams.