OpenAI's wet-lab experiment shows the prompt scaffold is doing real work

What was actually run at the bench

OpenAI worked with Red Queen Bio, a biosecurity start-up, to test whether GPT-5 could improve a real molecular biology protocol through experimental feedback. The task: optimize a two-piece cloning reaction joining a GFP gene into the pUC19 plasmid, starting from New England Biolabs' HiFi assembly, itself based on Gibson assembly.

Over five rounds of enzymatic optimization (44 reactions) plus a one-shot transformation screen (13 protocols), GPT-5 improved end-to-end cloning efficiency 79-fold—79x more sequence-verified clones per fixed amount of input DNA. The enzymatic change contributed 2.6-fold and the transformation change 36-fold; combined, they were additive to 79x.

The headline chemistry is a genuinely unfamiliar combination: GPT-5 proposed adding E. coli recombinase RecA and phage T4 gp32 single-stranded DNA-binding protein, staged around a temperature cycle (50°C, then 37°C for the new proteins, then back to 50°C). OpenAI states that, to its knowledge, RecA and gp32 have not been functionally co-used in molecular biology methods.

The fixed-prompt loop is the interface being tested

The most deliberate design decision here is not the model but the harness around it. Prompting was standardized with no human input beyond clarifying questions. Each round, GPT-5 proposed a batch of 8-10 reactions; scientists ran them and uploaded colony counts; the best data fed the next round. The only human role was physical execution and data entry.

That choice let OpenAI attribute the novel mechanism to the model rather than to human steering. But the same fixed scaffold produced a specific limitation, which the write-up is candid about.

This scaffolding helped reveal the model's capacity to propose genuinely novel protocol changes independent of human guidance, but it also locked the system into exploration and limited its ability to maximize the performance of newly discovered ideas.Montana Labs

In other words, the interface biased the run toward discovery over refinement. OpenAI expects that advances in planning and task-horizon reasoning would let simple fixed prompts support both. The lesson for anyone building agent loops: the prompting frame is not neutral plumbing—it decided whether the system explored broadly or exploited a good idea, and it left both the enzymatic and transformation gains under-optimized.

A plain-English layer between model and robot arm

To raise throughput, Robot on Rails and Red Queen Bio built a system that ingests a natural-language cloning protocol and executes it. It stacks three parts: a human-to-robot LLM that converts plain English into robot actions, a vision system that localizes labware in real time, and a path planner that carries out each action.

Running the standard HiFi method and R8 (the top first-round AI-modified protocol) side by side, the robot reproduced the ranking: human-executed R8 showed a 2.39-fold improvement, robot-executed R8 reached 2.13-fold—89% of human performance. But absolute colony counts from the robot were roughly ten-fold lower than manual execution.

That gap is instructive. A natural-language protocol translated cleanly enough to preserve relative rankings, but the last mile—liquid-handling precision, temperature calibration, the tacit nuances of manual cell handling—still cost an order of magnitude in yield. Translating intent into English is one interface problem; translating English into reliable physical action is a separate, harder one.

When the interface is the safety surface

OpenAI framed this as a biosecurity evaluation, run in a tightly controlled setting: a benign experimental system, a limited task scope, and explicit prompt-level prohibitions—for instance, the prompt forbade the use of cell extracts. The evaluation ties into the Preparedness Framework and its plans for model- and system-level safeguards.

The specific implication is that in embodied scientific agents, the frontend and the safeguard are the same surface. The prompt that constrained the search space, the clarifying-question protocol, and the natural-language-to-robot translator are all points where behavior was both shaped and bounded. When a model's outputs become physical reactions at a bench, the interface layer stops being a convenience and becomes the place where capability and control are jointly decided.