OpenAI Tests GPT-5 on Real Wet Lab Biology Work

OpenAI has introduced a real-world evaluation framework that measures how AI can accelerate biological research in the wet lab. The company used GPT-5 to optimize a molecular cloning protocol, examining both the promise and the risks of AI-assisted experimentation.

The work marks a shift from standard benchmark testing toward practical laboratory performance. Instead of scoring a model on abstract questions, OpenAI measures what the system can actually do when paired with concrete experimental protocols of the kind biologists run every day.

Molecular cloning — assembling and copying DNA fragments — is a routine but detail-heavy task in biology. Protocols involve precise reagent concentrations, incubation times, temperature steps, and sequencing-dependent branching decisions. Optimizing such a protocol is exactly the kind of iterative, constraint-driven reasoning that tests whether a large language model can function as a working research assistant rather than a chat interface.

OpenAI says the framework explores two sides of the same capability. On one side, AI acceleration of wet lab work could compress research cycles that currently take weeks. On the other, the same competence that helps a legitimate researcher streamline an experiment could lower barriers for misuse. The company frames its evaluation as a way to observe both outcomes directly rather than speculate about them.

The stakes sit at the intersection of two debates. The first is scientific productivity: if models like GPT-5 can reliably improve experimental protocols, AI moves from summarizing literature to generating laboratory value. The second is biosecurity policy: regulators and researchers have repeatedly flagged AI-assisted experimentation as a category of risk that requires measurement before deployment at scale.

By publishing a real-world evaluation rather than a static benchmark, OpenAI gives other labs a template for testing biological capability under conditions that resemble actual research use. The framework itself — not just GPT-5's score on it — may become the more consequential output, since standardized measurement is the precondition for any serious governance of AI in biology.

OpenAI has not announced what comes after the GPT-5 cloning experiment, but the direction is clear: the company intends to keep measuring AI capability where it matters most, in the lab rather than on the leaderboard.