OpenAI reports that GPT-5 improved the efficiency of a standard molecular cloning protocol by 79× in a controlled wet-lab study with Red Queen Bio. The model proposed novel changes (including an enzyme-assisted assembly approach) and a separate transformation tweak; humans executed the experiments and validated results across replicates. Early but significant.

What happened?

OpenAI worked with Red Queen Bio to test whether an advanced model could meaningfully improve a real experiment. GPT-5 proposed protocol changes; scientists ran the experiments and fed results back; the system iterated. Outcome: 79× more sequence-verified clones from the same DNA input versus the baseline method — the study’s definition of “efficiency”.

Why it matters: cloning is a core building block across protein engineering, genetic screens and strain engineering, so higher yield per input can shorten cycles and lower cost across a lot of everyday biology.

What actually changed?

• New assembly mechanism: GPT-5 suggested an enzyme-assisted variation that adds two helper proteins (RecA and gp32) to improve how DNA ends find and pair — a step that limits many homology-based assemblies. This alone improved efficiency in the study.

• A separate transformation tweak: It also proposed a handling change during transformation that increased the number of colonies obtained. Together, the assembly change and the transformation change delivered the 79× end-to-end improvement in the study’s validation runs.

Note: The team emphasises this was done in a benign system, with tight safety controls, and that results are early and system-specific — promising, but not a general guarantee.

How “79× efficiency” was measured

Efficiency here means sequence-verified clones recovered per fixed amount of input DNA compared with the baseline cloning protocol. OpenAI reports validation across independent replicates (n=3) for the top candidates.

What this doesn’t mean

• It doesn’t mean unsupervised AI running a free-form lab. Humans executed the experiments; the model proposed and iterated.

• It doesn’t mean the improvement applies to every organism, vector, insert or workflow. The team notes the gains were specific to their set-up and that broader generalisation requires more work.

• It doesn’t remove safety concerns. The work followed a preparedness framework and a constrained, benign system to manage biosecurity risk.

What’s genuinely new

• Novel, mechanistically-grounded idea: the RecA/gp32 approach formalises a “helper-assisted pairing” step within a Gibson-style workflow — interesting because Gibson has been a one-tube, one-temperature staple since 2009.

• AI–lab loop evidence: fixed prompting, no human steering in the proposal stage, yet still yielded a new mechanism plus a practical transformation improvement.

• Early robotics signal: the team also trialled a general-purpose lab robot that ran AI-generated protocols; relative performance tracked human-run experiments, albeit with lower absolute yields (areas for calibration remain).

Practical implications for R&D leaders

• Expect faster design–make–test cycles where benign “model systems” are used for method development, then adapted by domain experts.

• Plan governance: treat AI as a proposal engine inside a safety-first framework (risk review, change control, audit trail).

• Investment thesis: if even a fraction of these gains generalise, cost/time per cloning step could drop materially — compounding across library construction and screening programmes. Independent coverage echoes this potential but cautions against hype.

FAQs

How did GPT-5 achieve 79×?
By combining a new assembly mechanism (with helper proteins to improve pairing) and a transformation-stage change, validated against a standard baseline; the metric was verified clones per fixed input DNA. OpenAI

Was the AI running the lab?
No. GPT-5 proposed and iterated; trained scientists executed and uploaded results. The study deliberately used fixed prompts to measure the model’s own contributions. OpenAI

Is this safe?
The experiments were done in a benign system under tight controls and framed within OpenAI’s preparedness approach. The authors explicitly highlight biosecurity considerations. OpenAI

Will the same gains appear in my lab?
Not guaranteed. The team stresses system-specific results and early-stage status; replication and broader benchmarking are needed. Independent journalists also note the field’s history of over-claiming — healthy scepticism applies. OpenAI

What was the baseline?
A Gibson-style assembly workflow — widely used for joining DNA fragments. The study positions its changes relative to that baseline. OpenAI