Ginkgo and OpenAI reportedly ran 36,000 autonomous biology experiments with GPT-5

What happened

A report circulating on dev.to's home feed claims that in February 2026, Ginkgo Bioworks connected OpenAI's GPT-5 to its cloud laboratory and let it run a biology campaign at unusual scale: 36,000 experimental conditions spread across six iterative cycles.

The dev.to post, which credits miguelguerrero.eu as its origin, describes the arrangement in simple terms. GPT-5 planned a batch of experimental conditions; Ginkgo's automated lab physically carried them out; the results went back to the model, which used them to plan the next round. Repeat six times.

Ginkgo is a synthetic biology company whose business is built around heavily automated, foundry-style laboratories, where experiments are specified programmatically and executed by robotics. That background matters for plausibility: a facility that already accepts structured experiment definitions can, in principle, accept them from a model as readily as from a human engineer.

How the loop reportedly worked

The setup is what researchers call closed-loop experimentation. Instead of a model acting purely as an analyst, scoring candidates or predicting outcomes while humans handle the bench work, the algorithm sits inside the experiment itself. It chooses what to test, observes real-world results, and revises its strategy before the next test.

The report does not describe the biological target, the assays involved, or how a condition was counted, so the scientific weight of the run is hard to judge. In general usage, a condition is one distinct combination of experimental variables, such as a strain in a given medium at a given temperature. Taken as an average, 36,000 conditions over six cycles works out to about 6,000 per cycle, which points to high-throughput batched design rather than 36,000 sequential decisions. That is consistent with how automated foundries already operate.

A big claim on thin sourcing

Some caution is warranted. The dev.to item is a short aggregation post rather than original reporting: it names miguelguerrero.eu as its underlying source, carries a disclosure noting it was produced with AI assistance and checked afterward, and includes no quotes, no statements from Ginkgo or OpenAI, and no pointer to a primary announcement or publication. Until either company documents the work, the headline numbers should be treated as a secondhand claim rather than a verified result.

Why it matters

If the figures hold up, the milestone is less about the raw count than the role reversal. AI has been a powerful analytical layer in biology for years, from protein structure prediction to literature mining to candidate ranking, but the empirical loop, the part where a hypothesis meets physical lab work, has stayed human. A model that designs experiments, observes real outcomes and redesigns accordingly is doing empirical research rather than analysis.

That compression matters because design-test-learn cycles in biology are usually bounded by lab throughput and human planning time. An agent that shrinks each cycle from weeks to days changes which optimization problems are worth attempting, whether in strain engineering, enzyme tuning or media development.

It also surfaces the questions that follow any autonomous system in a wet lab: what biosafety review applies to experiments a model selects on its own, how thousands of autonomous design decisions can be audited after the fact, and where the guardrails sit. The report projects laboratories where AI handles iterative experimentation while people set direction rather than individual steps, and that is precisely the scenario those oversight questions will need to answer.