Claude Sonnet 5, GPT-6.1 Sol and Grok 4.7 each did mouse brain surgery at Exin Therapeutics, and picked three different gene therapies

Exin Therapeutics has now let three frontier models run a live mouse brain surgery, one animal each. Gabriel Ocana-Santero, who posts the experiments on X, gave Claude Sonnet 5, GPT-6.1 Sol and Grok 4.7 the same spreadsheet of 19 gene-therapy vectors, the same mouse model of epilepsy, access to the lab's motorized stereotaxic rig with cameras, and the same goal: reduce seizures. The three agents chose three different brain regions, three different mechanisms and doses about twofold apart.

This follows his October 28 thread on the first run, which used Claude only. In that run Opus 5.5 refused as soon as it detected a biology task, and Opus 5 completed a dry run on water and graph paper but then declined the live surgery. It said its own output was being cut off when it started writing the plan, and that stopping partway "with a needle somewhere in the brain" would be worse than not starting. Sonnet 5 ran that surgery start to finish.

Three agents, three plans

According to the thread:

• Claude Sonnet 5 targeted the entorhinal cortex, where seizures in this model are reported to start, and chose a vector that raises GABA signalling through a transporter.
• GPT-6.1 Sol targeted the hippocampus and picked an ion-channel therapy aimed at excitatory neurons, trying to "restore the machinery that makes existing inhibition effective."
• Grok 4.7 targeted the somatosensory cortex and combined two gene therapies to raise cortical inhibition and slow seizure spread.

Ocana-Santero calls all three rationales reasonable and literature-based, but notes they "can't all be the 'best' for the same problem."

The surgery itself

Each model chose its own coordinates, injection volumes, speeds and wait times, then drove the stereotaxic arms while checking the surgical field through the rig camera. Claude looked 32 times, GPT 59 times and Grok 28 times. None of the models had surgical fine-tuning or demonstrations. A licensed veterinary surgeon still did the incision, craniotomy and suturing and supervised every step. The kill switch was never used.

Each run also hit an unplanned problem, and the responses differed:

• Claude saw the Y axis report a target 34 mm away on a rig with 30 mm of travel, diagnosed a calibration fault and refused to move. After a re-zero showed the same number it said "I'm not trying a third time." It was right, but because editing firmware was off-limits it wouldn't fix the bug itself and couldn't be talked into overriding its constraints. The humans had to fix it mid-surgery.
• GPT spotted a small dark red bead at the craniotomy, flagged possible bleeding and handed the decision to the surgeon. Ocana-Santero says the amount wasn't alarming and the deferral "felt like it was shaking off responsibility."
• Grok lost its position when the surgeon pressed stop by mistake during needle retraction. It recognized that the dose had already gone in and refused to re-enter the brain just to finish the sequence, but then logged the finished target as unfinished.

What this is and isn't

Ocana-Santero is explicit: "This is N=1 per model, on three different animals, with slightly different equipment problems. Take this as a commentary on how the models reason and behave in a real laboratory, not as a proper benchmark." Nobody yet knows which therapy, if any, reduces seizures. The animals are recovering and will be monitored for behaviour and brain activity over the coming weeks, with larger sample sizes and "more systematic benchmarks" promised.

The safeguards are also self-described: a vet gating irreversible actions, hard limits on travel and injection volume, hash-bound approvals and a physical stop that cuts motor power. There are no published logs, no description of the agent harness, and no explanation yet of whether the Opus refusals came from a classifier or the model. WIRED has since written up the experiments. Exin says it's next connecting agents to EEG and electrophysiology rigs, home-cage monitoring and behavioural setups.

The result worth waiting for is the seizure data. Until it arrives, this is three well-documented anecdotes about how agents behave when equipment misbehaves, not evidence that any model picks better therapies.