OpenAI and Ginkgo put GPT-5 in a closed loop with a robotic wet lab

What the loop actually did

OpenAI paired GPT-5 with Ginkgo Bioworks' cloud laboratory — an automated wet lab operated remotely through software, where robots run experiments and return data. GPT-5 designed batches of reactions in standard 384-well plate format, the lab executed them, and the results flowed back to the model to shape the next round.

They ran that cycle six times. Across the full run the system executed more than 36,000 unique cell-free protein synthesis (CFPS) reaction compositions across 580 automated plates. After being given a computer, a web browser, and access to relevant papers, GPT-5 took three rounds and two months to set a new state of the art: a 40% reduction in protein production cost against the prior best baseline, and a 57% improvement in reagent cost.

The task itself is a good fit for this treatment. CFPS makes proteins without growing living cells, running the protein-making machinery in a controlled mixture of DNA template, cell lysate, and many interacting biochemical components. It is hard to optimize by intuition because small changes matter and the direction of their effect isn't obvious.

The validation step that keeps the model honest

The detail worth dwelling on is the guardrail. Before any experiment ran, OpenAI added strict programmatic validation enforcing that AI-designed experiments were physically executable on the automation platform.

It prevented "paper experiments" that look plausible in text but can't be carried out in a robotic workflow.Montana Labs

This is the part of the system doing real work. A language model can generate a chemically reasonable-sounding protocol that a pipetting robot simply cannot perform. By making executability a hard precondition rather than something discovered after a failed run, the setup ties the model's fluency to the mechanical constraints of the lab it's actually driving.

Where the model diverged from human habit

The findings are more specific than "AI found better recipes." GPT-5 identified low-cost compositions that humans had not previously tested in this configuration, and many of them held up under conditions that trip up bench-scientists: high-throughput plate reactions run at low volume, with lower oxygenation and different mixing than test tubes.

OpenAI reports that small changes in buffering, energy regeneration components, and polyamines had an outsized impact relative to their cost — parameters people don't usually reach for first. At scale, those become testable hypotheses rather than background assumptions.

And the cost structure steered the strategy. In this CFPS system, costs are dominated by lysate and DNA, which makes yield the highest-leverage lever: boosting protein output per unit of expensive input moves the cost number more than trimming cheaper reagents. That reframing — optimizing for output per costly input rather than for cheaper ingredients — is the kind of reasoning the throughput made legible.

The narrow claim underneath the automation story

OpenAI is careful about scope, and applied teams should be too. The result was demonstrated on one protein, sfGFP, and one CFPS system; generalization to other proteins and other systems is unproven. Some gains may be sensitive to oxygenation and reaction geometry that vary across scales. And human oversight was still required for protocol improvements and reagent handling.

So the specific implication of this announcement is not that a model discovered new biology unaided. It is that iteration throughput — 36,000 reactions on robots, gated by executability checks — is what turned a model's proposals into a measurable frontier shift, on a task where cost matters precisely because autonomous labs can run thousands of reactions where a human team runs dozens. The bottleneck this addresses is iteration speed, and that is a claim narrow enough to be useful and honest enough to build on.