GPT-5 breaks into the wet lab with a 79x cloning boost

AI flexing its muscles in math proofs and code is old news. But moving from pure abstractions to the chaotic reality of a wet lab is a whole different beast—molecules simply do not care how confident an LLM sounds.

OpenAI just revealed an experiment that bridges that gap. Partnering with biosecurity startup Red Queen Bio, the company put GPT-5 in charge of optimizing a fundamental molecular biology procedure: cloning a green fluorescent protein gene into a standard pUC19 plasmid. Running in an autonomous loop with fixed prompts and zero human guidance, the model came up with a protocol that boosted cloning efficiency by a massive 79x.

Detangling DNA with a novel enzyme duo

The team started with HiFi assembly, a proprietary system developed by New England Biolabs based on the classic Gibson assembly method introduced back in 2009. Gibson assembly is beloved because it is isothermal and happens in a single tube. But that simplicity also leaves performance on the table.

GPT-5 broke those constraints by proposing a new mechanism it dubbed RAPF-HiFi. It pulled in two specific proteins: RecA from E. coli and the phage T4 gene 32 single-stranded binding protein, known as gp32.

The mechanics are clever: gp32 behaves like a comb, smoothing out messy, folded single-stranded DNA ends. Then, RecA acts like a matchmaker, actively pairing the strands together. On top of that, GPT-5 engineered a three-step thermal schedule—running an initial reaction at 50°C, cooling it to 37°C so the helper enzymes could work, and bumping it back to 50°C to release them and finish assembly. That alone provided a verified 2.5x jump.

Challenging textbook rules on cell handling

The bigger shock came during the transformation step, where engineered DNA is put into bacteria. Molecular biology lore treats chemically competent cells as extremely fragile glass cannons that must be handled gently.

GPT-5 ignored the superstition. In a one-shot optimization test, it suggested pelleting the cells in a centrifuge at 4°C, dumping half the liquid volume, and gently resuspending them to double cell density and boost molecular collisions. The supposedly fragile cells handled it fine, and the simple trick produced a more than 30-fold surge in transformation efficiency. Multiplying the two interventions together delivered the final 79x leap.

Throughout the multi-round process, human scientists merely served as hands: pipetting reagents, counting colonies, and logging data back into the system without offering any advice.

Why it matters

This experiment is a major milestone because it drags frontier models out of text generation and puts them squarely into empirical discovery. For researchers in protein engineering, synthetic biology, and genetic screening, a 79-fold yield improvement means vastly larger libraries and substantially lower reagent costs without needing exotic new machinery.

We think the setup of this test is particularly smart. By locking the model into fixed prompts without human steering, OpenAI proved that the mechanistic leap was generated by machine reasoning rather than human prompt tweaking. The downside, as the researchers acknowledged, is that the rigid framework over-indexed on exploration, meaning the model likely left even greater gains on the table by not fully optimizing its best hits.

Looking at the bigger picture, this is a watershed moment for lab automation, but it also lights a fire under biosecurity policy. If an AI can independently overhaul molecular biology protocols to make them exponentially more potent, the barrier to synthesizing both life-saving cures and dangerous biological agents is dropping fast. The wet lab assistant has officially arrived, and biology will have to adapt quickly.