Article · AI translation
On November 8, Beijing time, according to earlier reporting by WIRED and the latest disclosures from biotechnology company Exin Therapeutics, OpenAI's AI model GPT-6.1 Sol has been used in brain surgery experiments on live mice and improved its localization performance through surgeons' feedback over 3 consecutive procedures. By the third procedure, the time it took to identify key anatomical landmarks had fallen from 169 seconds to 86 seconds, and its localization error from 0.6 millimeters to 0.3 millimeters.
Exin Therapeutics co-founder and CEO Gabriel Ocana-Santero announced the results on the social platform X on November 5. The research team had GPT-6.1 Sol perform anatomical landmark localization tasks in 3 consecutive brain surgeries on live mice. After each procedure, a professional surgeon corrected the positions identified by the AI, while the model saved the relevant images, measurements and correction records for reference in the next procedure.
The experiment focused mainly on 2 key anatomical landmarks on the surface of the mouse skull: Bregma and Lambda. These 2 locations are important reference points in stereotactic brain surgery. Researchers use them to determine brain region coordinates so they can accurately guide an injection needle or other experimental instruments to the target location.
The research team said that over 3 rounds of procedures, the number of manual corrections needed for GPT-6.1 Sol fell from 4 to 2, and localization error decreased by 50%. Notably, this improvement was achieved without retraining the model or updating its parameters. Instead, it used experience accumulated during earlier procedures and expert feedback to adjust its judgments in subsequent tasks, a practical application of in-context learning.
This was not the first time Exin Therapeutics had involved AI in brain surgery on live subjects. Earlier, the research team separately involved Anthropic's Claude Sonnet 5, OpenAI's GPT-6.1 Sol and xAI's Grok 4.7 in gene therapy experiments on mouse brains.
The 3 AI models were given the same 19 candidate gene therapy options. Each was tasked with choosing a treatment for mice with epilepsy and administering injections using motorized stereotactic brain equipment. The results showed that the 3 models chose different treatment mechanisms and target brain regions: Claude chose the entorhinal cortex, GPT-6.1 Sol chose the hippocampus, and Grok chose the somatosensory cortex.
During the experiments, the AI models could observe the surgical area through cameras, plan injection coordinates, dose, speed and waiting time on their own, and control the equipment to carry out the corresponding operations. In one experiment, GPT-6.1 Sol checked the camera view 59 times. It also spotted a small dark red dot at the craniotomy site, suspected possible bleeding and referred its assessment to the surgeon. The researchers later said the phenomenon was not abnormal bleeding for this type of surgery.
However, the AI did not perform these surgeries entirely on its own. A licensed veterinarian remained responsible for procedures such as making incisions, opening the skull and suturing, and supervised the entire surgery. Irreversible critical steps required human approval. The experimental equipment also had limits on its range of motion and injection dose, as well as an emergency stop switch, to prevent harm from AI operating errors.
WIRED previously reported on the team's early experiment using Claude Sonnet 5. At the time, the researchers initially tried to have Claude Opus 5.5 carry out the task, but the model refused to advance the injection needle into the mouse's brain. The team then switched to Claude Sonnet 5 to complete the injection. Building on that work, the researchers further tested GPT-6.1 Sol and Grok 4.7 and conducted the latest experiment involving consecutive procedures.
The experiments are still at an early exploratory stage. The research lead acknowledged that a sample of 3 procedures is very limited and is not yet enough to show that AI can reliably master the relevant skills. Whether the treatments chosen by the different models are actually effective will also require subsequent observation of the mice's behavior and brain activity.
These preliminary experiments show the possibility of AI moving beyond analyzing research data and devising experimental plans to directly operating real experimental equipment. However, the findings have not yet undergone sufficient independent verification and peer review, and their application in human brain surgery remains a long way off.