OpenAI unveiled on February 5 a partnership with Ginkgo Bioworks around a “looped” system where GPT-5 designs experiments, pilots a cloud wet lab, controls robots, analyzes data and plans subsequent iterations. Immediate objective achieved: reduce the cost of cell-free protein synthesis (CFPS) by approximately 40%, with a reduction of 57% on the reagent side.
In this configuration, the model had access to the Internet, scientific literature and analysis tools. The protocol has been validated to guarantee that each experimental plan can be physically executed by the robots, avoiding purely theoretical dead ends. Testing included more than 36,000 unique formulations spread across 580 automated microplates.
Three iteration cycles were enough to surpass the previous best human benchmark. GPT-5 performed well in exploring high-dimensional parameter spaces, identifying low-cost combinations that teams had not previously tested.
Results and lessons
The new recipes are distinguished by a marked robustness in conditions of low oxygenation, typical of automated laboratories. The model also highlighted the effect of discrete levers, such as the adjustment of buffers and polyamines, allowing a significant gain in yield for minimal additional cost.
In total, the human–AI approach reduced overall CFPS costs by 40% and reagent costs by 57%. Beyond the quantified performance, the interest lies in the stability of the results in constrained environments and the capacity for rapid exploration of formulation areas little invested by human intuition.
Technical scope and implied limits
The direct AI–wet lab coupling, with robotic execution and intermediate quality control, confirms the interest of closed loops in experimental sciences. The scale of testing shows an increase in maturity of cloud/automation orchestration, but gains remain conditional on the validity of sensors, metrology and standardization of consumables. Generalization to other biosynthetic pathways will depend on the transposability of the learned parameters and batch constraints.
For the ecosystem, cheaper and more robust CFPS expands use cases in enzyme prototyping, rapid production of specific proteins, and field biology, with a likely ripple effect on cloud automation platforms and modular reagent providers. If model-driven iteration takes hold, the value will shift toward lab data quality, protocol traceability, and the ability to integrate reliable hardware loops rather than just computing power.
Source : It’s
