OpenAI Says GPT-5 Cut Protein Production Costs 40% in Robotic Lab

OpenAI reports that GPT-5 connected to Ginkgo Bioworks' cloud laboratory achieved a 40% reduction in protein production cost in cell-free protein synthesis, setting a new state of the art in low-cost CFPS after three rounds of experimentation and roughly two months of work.

The system ran more than 36,000 unique CFPS reaction compositions across 580 automated plates over six rounds of closed-loop experimentation. Alongside the 40% cost reduction, OpenAI reports a 57% improvement in the cost of reagents, including novel reaction compositions that hold up better under reaction conditions common in autonomous labs.

The results matter beyond a single benchmark. Biology has lagged math and physics in AI acceleration because progress runs through physical experiments that take time and money. Frontier models plugged into lab automation — proposing experiments, running them at scale, learning from results, and deciding next steps — target that bottleneck directly. In much of life science, OpenAI argues, iteration is the constraint, and autonomous labs exist to remove it.

How the loop worked

OpenAI paired GPT-5 with Ginkgo Bioworks' cloud laboratory, an automated wet lab operated remotely through software where robots execute experiments and return data. GPT-5 designed batches of experiments in a standard 384-well plate format. The lab executed them. The cloud laboratory pushed the data back to GPT-5, where the model analyzed outcomes, generated new hypotheses, and designed the next round. The cycle repeated six times.

The team added strict programmatic validation before any experiment ran. That validation enforced that AI-designed experiments were physically executable on the automation platform. It blocked "paper experiments" that look plausible in text but cannot be carried out in a robotic workflow.

The scale was the point. Single experiments in biology are noisy, and throughput and iteration are how researchers separate signal from random noise. Once GPT-5 had access to a computer, a web browser, and relevant papers, it proposed many plate-based reactions that outperformed the prior best baseline (Olsen et al., 2025) immediately.

Why CFPS matters

Cell-free protein synthesis makes proteins without growing living cells. Instead of putting DNA into cells and waiting, CFPS runs the protein-making machinery in a controlled mixture. That makes it a practical tool for rapid prototyping: scientists can run many experiments and measure results the same day.

Proteins carry a large share of what modern biology delivers. Many medicines are protein-based. Diagnostics and research assays depend on them. Industrial enzymes make chemical processes cleaner and more efficient. Proteins appear in laundry detergent. When protein production gets faster and cheaper, scientists can test more ideas sooner and cut the cost of turning early research into products.

The bottleneck is that CFPS is tricky to optimize and expensive at scale. The system involves complex, interacting ingredients: the DNA template encoding the target protein, the cell lysate containing cellular machinery, and a long list of biochemical components from energy sources to salts. Previous studies have applied machine learning to reduce production cost, but progress has been incremental because exploring the mixture space thoroughly is labor-intensive.

Cost dynamics compound the problem. Standard CFPS formulations and commercial kits are priced for human-paced work. An autonomous lab can run thousands of reactions in the time a human team runs dozens, making reagent cost the limiting factor.

What GPT-5 found

The improvements came from identifying combinations of components that work well together and hold up under the realities of high-throughput automation.

GPT-5 identified low-cost reaction compositions that humans had not previously tested in this configuration. CFPS has been studied for years, but the space of possible mixtures remains large, and proposing and executing thousands of combinations quickly surfaces workable regions that manual workflows miss.

High-throughput, plate-based experiments also differ from manual bench-top ones. Oxygenation can be lower; mixing and geometry differ. Most CFPS reactions produce much more protein in test tubes than in microtiter plates, because larger scales bring more oxygen availability and better mixing. GPT-5 proposed many reagent combinations that performed well under these constraints, including many that are more robust in the low-oxygen conditions common in automated labs.

Small changes had outsized effects. Adjustments to buffering, energy regeneration components, and polyamines delivered gains far beyond their cost. These are not the parameters researchers typically reach for first, but at high throughput they become testable hypotheses rather than background assumptions.

The cost structure itself shaped the strategy. Costs in CFPS are now dominated by lysate and DNA, which makes yield the highest-leverage variable. Boosting protein output per unit of expensive input delivers meaningful cost progress before chasing marginal savings elsewhere.

Limitations

OpenAI is explicit about the boundaries of the result. The experiments demonstrated performance on one protein, sfGFP, and one CFPS system. Generalization to other proteins and other CFPS systems still needs to be shown.

Oxygenation and reaction geometry strongly affect yields and vary across scales, so some improvements may be sensitive to these conditions. Mapping those sensitivities is part of what comes next.

Human oversight was required for protocol improvements and reagent handling. The system can design and interpret experiments, but lab work still involves practical details that need experienced operators.

What's next

OpenAI plans to apply lab-in-the-loop optimization to other biological workflows where faster iteration can unlock progress. The company frames autonomous labs as complementary to models: models generate designs, but biology still requires testing and iteration, and closing the loop between the two is what turns promising ideas into working results.

The company also flags biosecurity stakes. These results show that models can reason in the wet lab to improve protocols, which "may have implications for biosecurity that we assess and mitigate through our Preparedness Framework." OpenAI says it is committed to building safeguards at the model and system level and developing evaluations to track current risk levels.