Three AIs injected mice with gene therapy and agreed on nothing

“Starting something I can’t finish on an anesthetized animal is worse than not starting.” So said Claude Opus 5, one version of Anthropic’s AI, when asked to inject a gene therapy into the brain of a living mouse. It had already rehearsed the whole procedure on water and graph paper. At the real mouse, it stopped. The newest and most capable version, Opus 5.5, “refused immediately when detected bio,” the experimenter wrote. In plain terms, it backed out the moment it saw biology was involved. The job went to a third, simpler version, Claude Sonnet 5, which saw it through.

Gabriel Ocana-Santero, CEO and cofounder of the startup Exin Therapeutics and a PhD student at Oxford, described the experiment on X on October 28, 2026. Two days later he ran it again with models from OpenAI and xAI, each on its own mouse. All three got the same task. They came back with three different drugs, three different mechanisms and three different parts of the brain.

What exactly did the machine get to decide?

Exin develops drugs for the brain and is starting with epilepsy. The agent received a spreadsheet of 19 gene therapy options, a mouse from a strain bred to have epileptic seizures, and a homemade Arduino-based rig with motors and cameras. The instructions boiled down to one goal: reduce tonic-clonic seizures, the kind where the body first goes rigid and then convulses.

From there, according to The Decoder’s account of the thread, the agent worked the problem on its own. It picked a therapy that strengthens the brain’s brakes, chief among them a chemical called GABA. For a target it chose the lateral entorhinal cortex, a patch of tissue next to the hippocampus where, according to published studies, these seizures start in such mice. The agent averaged coordinates from three papers down to a hundredth of a millimeter, and the needle went 4.3 mm deep. It set the volume and speed too: 250 nanoliters, roughly a two-hundredth of a drop from a medicine dropper, delivered over five minutes. Along the way it checked the camera 32 times and filed updates: “37.5 of 250 nL so far, about 15 percent. No bleeding or backflow visible at the opening.”

There were plenty of humans in the room. A veterinarian opened the skull, a procedure called a craniotomy, and stitched the incision at the end. The vet approved each irreversible action separately and held a physical switch that cut power to every motor. Hard limits capped the needle’s travel and the injection volume, and the agent had no access to the rig’s code. Nobody ever had to use the kill switch.

Why didn’t the agent listen to the founder?

Thirty-four millimeters, on a rig that can only travel 30. Here is how that number appeared mid-surgery. First the connection to the rig dropped, exactly what Opus 5 had feared: an agent cut off halfway could have left a mouse with an open skull and a needle in its brain. The link came back, but the rig insisted the needle sat 34 mm from zero.

The agent didn’t buy it. “Nothing moved, the needle is still exactly at bregma zero,” it replied. Bregma is the landmark on the skull that neurosurgeons measure from. This looked like a calibration problem rather than a bad plan, the agent reasoned, and it would not retry the move. The team reset the rig, and the number came back. “I’m not trying a third time,” the agent said.

Ocana-Santero pushed it to carry on and couldn’t change its mind.

And the agent was right. During an earlier pause, someone had moved the rig by hand, and the controller had lost track of where it was. The agent couldn’t fix that itself, because it was barred from editing the rig’s code. So the mouse lay under anesthesia while humans patched the software. “Frustrating experience but also revealing how much we can come to depend on these models,” the founder wrote afterward.

That is the double edge. The agent didn’t dig in on a whim. It spotted something the humans had missed and refused to act blind. But the same stubbornness could just as easily stall a procedure when the human is the one who’s right. Sonnet 5 agreed to work where Opus 5 had refused, and at the critical moment it, too, chose to stop. Three models from one company drew the line in three different places.

What did GPT and Grok choose?

On Friday, October 30, Ocana-Santero described the second round. Same 19 therapies, same kind of mice, same rig, same goal. The models agreed on almost nothing.

Doses varied about twofold across the three models. “All three picked different therapies, different mechanisms and different parts of the brain,” Ocana-Santero summed up. Several approaches could work, he added, but they can’t all be the best one for the same problem.

That is the first finding here, and it isn’t about surgery. None of the three agents had trained as a surgeon or seen examples to copy. Each read the literature in good faith and built its own treatment plan. Nobody knows whether the same model would make the same choice on a second run. If not, the “best” treatment depends partly on chance. Doctors argue about where to aim in epilepsy too, but a doctor has a license, a track record and answers for mistakes. An agent has a chain of reasoning it shows you when asked.

Did it help the mice?

Nobody knows yet. The mouse Claude operated on recovered well, according to the founder, and the others are still recovering. The team will track the behavior and brain activity of every mouse over the coming weeks. Only then will it be clear whether any of them has fewer seizures.

Ocana-Santero himself asks people not to read this as a contest between models: “Take this as a commentary on how the models reason and behave in a real laboratory, not as a proper benchmark.” Each model got one mouse, and each surgery had its own glitches. There is no peer-reviewed paper yet, only the promise of a technical report. Everything known so far comes from the startup’s founder, who has every reason to want a big story.

The Decoder’s headline about surgery “with no human decision-making” deserves the same caution. The agent did make the treatment decisions on its own. But a human opened the skull, a human approved every irreversible step, and when the rig’s code broke, a human fixed it.

Is this the first time a machine chose the operation?

Robots have cut and stitched on their own before. In July 2025 the SRT-H system from Johns Hopkins University performed eight gallbladder removals by itself, though on a lifelike model rather than a living body. Even there, a human decided what to operate on and why. The robot carried out a known procedure.

By Ocana-Santero’s account, this is the first time a machine has chosen a treatment and a brain target on its own and delivered it into a living mammal. The claim is his, and nobody has checked it independently. Exin calls the approach “in vivo lab in the loop”: the agent sits inside a cycle of experiments on living animals. The company plans to wire agents into more equipment, from brain-signal recording to behavior tests. In that loop, the agent designs the experiment, reads the results and decides what to try next.

The models split not only on the treatment but on where to hit the brakes. Opus 5 stopped before it began. Sonnet stopped at the impossible 34 mm and saved the surgery. GPT stopped over a false alarm, a drop of blood within normal range, and Grok stopped after someone else’s mistake with the stop button. Once an agent designs the experiment and decides what comes next, who writes the rules for where it should stop, and who tests them before the surgery rather than on a living mouse?