Exclusive: GPT-5 demonstrates ability to do novel lab work
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Sarah Grillo/Axios
GPT-5 has for the first time demonstrated it can do the kind of lab work that opens up a pathway for AI to take a bigger role in scientific experiments, OpenAI shared first with Axios.
Why it matters: The work shows how GPT-5 — and other AI models with similar capabilities — can speed up research, reduce costs and help human scientists make real-world discoveries.
The big picture: Major AI gains in biology have come more slowly than fields like math or physics, since progress in biology depends on real-world lab work.
- Some AI skeptics have pointed to a lack of progress in science — if AI's so great, why hasn't it cured cancer yet? — as a "proof" the technology is over-hyped.
What they did: OpenAI worked with a biosecurity startup — Red Queen Bio —to build a framework that tests how models work in the "wet lab."
- Scientists use wet labs to handle liquids, chemicals, biological samples and other "wet" hazards, as opposed to dry labs that focus on computing and data analysis.
- In the lab, GPT-5 suggested improvements to research protocols; human scientists carried out the protocols and then gave GPT-5 the results.
- Based on those results, GPT-5 proposed new protocols and then the researchers and GPT-5 kept iterating.
What they found: GPT-5 optimized the efficiency of a standard molecular cloning protocol by 79x.
- "We saw a novel optimization gain, which was really exciting," Miles Wang, a member of the technical staff at OpenAI, tells Axios.
- Cloning is a foundational tool in molecular biology, and even small efficiency gains can ripple across biotechnology.
- Going into the project, Nikolai Eroshenko, chief scientist at Red Queen Bio, was unsure whether GPT-5 was going to be able to make any novel discoveries, or if it was just going to pull from published research.
- "It went meaningfully beyond that," Eroshenko tells Axios. He says GPT-5 took known molecular biology concepts and integrated them into this protocol, showing "some glimpses of creativity."
Of note: OpenAI used a "benign experimental system" and worked in a tightly controlled setting to prevent biosecurity risk with the experiments.
Between the lines: AI model benchmarks are notoriously limited.
- They look a lot like a student taking a test, says Wang.
- Together Wang and Eroshenko set out to test how GPT-5 would interact in this scenario. Like human scientists, Wang says, sometimes the model would propose crazy moon shot ideas with high risk and high reward, and other times it was more reserved.
Reality check: The AI industry — partly propelled by hype, relentless competition and a race against China — has gotten in trouble for misrepresenting scientific achievements.
- Wang was careful not to overstate the results. "It's not a foundational breakthrough in molecular biology. But I think it's accurate to call it a novel improvement, because it hasn't been done before."
What's next: Researchers say these findings are early, but point to a future where AI-augmented experimentation becomes routine, helping scientists translate insights into real-world impact faster than today's trial-and-error workflows.
