Ginkgo Bioworks Turned Its Lab Over to GPT-5
They claim that protein synthesis costs fell 40% after 36,000 automated experiments
Ginkgo Bioworks and OpenAI have published a preprint describing an autonomous laboratory system that uses a large language model to design and run biological experiments.
The work focuses on cell-free protein synthesis, a widely used laboratory method that is expensive and difficult to optimize.
According to the preprint and a company release, the system ran roughly 36,000 experimental conditions across six iterative cycles.
The goal was to reduce the cost of producing a standard benchmark protein, superfolder green fluorescent protein.
Ginkgo reports the system achieved reaction component costs of $422 per gram, compared with a previously published benchmark of $698 per gram under similar conditions.
The announcement drew attention in financial markets.
Ginkgo’s stock rose following the release, with market coverage pointing to investor interest in whether autonomous laboratories and AI-driven optimization could support future revenue through reagents, contract research services, or platform offerings.
The work is presented as an example of a language model operating as an experimental agent in biotech. Still, the scientific preprint that has not yet undergone peer review.
Letting AI Plan The Experiments
The preprint describes a closed-loop system connecting OpenAI’s GPT-5 reasoning model to Ginkgo’s cloud laboratory.
The model proposed experimental designs, selected reagent combinations, analyzed results from previous runs, and used those results to decide what to test next. Experiments were executed using Ginkgo’s automated lab infrastructure, including its reconfigurable automation carts and scheduling software.
The system ran more than 580 multi-well plates and generated nearly 150,000 experimental data points over the course of six months. Human involvement was limited but not eliminated.
According to the authors, people prepared reagents, loaded and unloaded plates, and monitored system performance, while experimental design, iteration, data interpretation, and hypothesis generation were handled by the model.
To prevent invalid or impractical experiments, each proposed design was checked against a predefined validation schema before execution.
The checks included plate layout, controls, replication, reagent availability, and volume constraints, and only experiments that passed validation were allowed to run.
“We expect more and more experiments to be run on autonomous labs where reagent and consumables costs dominate the cost of an experiment,” said Reshma Shetty, Ginkgo’s co-founder.
“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,” said Joy Jiao, OpenAI’s life sciences research lead.
Ginkgo said it plans to commercialize parts of the work. The company is selling the optimized reaction mix through its reagent store and has said it will release the experimental validation model as open-source software.
Protein Synthesis Costs Drop
The reported cost reduction applies to a specific case. The experiments focused on one protein and one cell-free protein synthesis system. The results do not claim lower costs across other proteins, production methods, or laboratory environments.
The comparison to “state of the art” is contextual. Published protein synthesis costs vary based on experimental design, scale, reagent sourcing, and performance targets.
The benchmark cited in the preprint reflects prior published work under comparable experimental conditions, not a universal industry baseline.
The paper provides detailed methods, experimental counts, and system architecture, allowing independent technical review. At the same time, some operational details remain limited in public materials.
While the authors describe automated validation checks, they do not fully disclose governance thresholds, failure-handling procedures, or escalation rules if the system proposes unexpected experimental paths.
The work neatly falls into a broader trend across biotechnology and pharmaceutical research, where companies are tying AI and automation to faster iteration and changes in R&D economics.
Recursion Pharmaceuticals has emphasized large-scale computing and automation as central to its discovery strategy, and its infrastructure announcements have previously coincided with stock price movements. Insilico Medicine has reported progress on AI-designed drug candidates and positioned AI as a way to shorten early discovery timelines.
Large pharmaceutical companies have taken similar steps through acquisitions and partnerships. AstraZeneca, for example, has acquired AI-focused companies to integrate machine learning more deeply into oncology research.
What distinguishes the Ginkgo-OpenAI work is the combination of experimental scale, methodological documentation, and immediate commercialization. The preprint describes how the system operated in detail rather than at a conceptual level, and the reagent offering turns the result into a product that can be independently evaluated.
At the same time, the findings remain limited to a single experimental domain, and the paper does not claim that similar gains will extend to other biological systems.