- Published on
OpenAI GPT-6.1 Sol participates in live mouse brain surgery: three consecutive operations, positioning error halved
- Authors

- Name
- DeepHorus
On November 8, Beijing time, according to previous reporting by WIRED and the latest disclosure from biotech company Exin Therapeutics, OpenAI's AI model GPT-6.1 Sol has participated in live mouse brain surgery experiments and improved its localization performance across three consecutive operations with the help of surgeon feedback. By the third operation, the time it took to identify key anatomical landmarks had shortened from 169 seconds to 86 seconds, and the localization error had decreased from 0.6 millimeters to 0.3 millimeters. Gabriel Ocaña-Santoro, co-founder and CEO of Exin Therapeutics, announced the results on the social platform X on November 5.
The experiment focused on Bregma (the anterior fontanelle) and Lambda (the posterior fontanelle) on the surface of the mouse skull. These two landmarks are important reference points for stereotaxic brain surgery. Researchers need to use them to determine coordinates for brain regions and guide instruments such as injection needles to target positions. After each operation, a professional surgeon corrected the positions identified by the AI, and the model saved the images, measurement data, and correction records for reference in the next operation.
After three rounds of operations, the manual corrections required for GPT-6.1 Sol decreased from four to two, and the localization error fell by 50%. The research team said that this progress was not achieved by retraining the model or updating parameters, but by using the experience accumulated from previous operations and expert feedback to adjust its judgments, which is a practical application of in-context learning. However, the experimental sample size remains very limited.
Previously, Exin Therapeutics also had Claude Sonnet 5, GPT-6.1 Sol, and Grok 4.7 participate in brain gene therapy experiments in mice with epilepsy. The three models were presented with the same 19 candidate gene therapy treatment plans and needed to choose a treatment approach and perform injections using motorized stereotaxic equipment. The results showed that Claude chose the entorhinal cortex, GPT-6.1 Sol chose the hippocampus, and Grok chose the somatosensory cortex. The team had tried to have Claude Opus 5.5 perform the task, but the model refused to advance the injection needle into the mouse brain, and they subsequently switched to Claude Sonnet 5 to complete the injection.
During the experiments, the AI could observe the surgical area through a camera, independently plan injection coordinates, dose, speed, and waiting time, and control the equipment to execute the procedure. GPT-6.1 Sol checked the camera images 59 times and also noticed a small dark red spot at the craniotomy site and suspected bleeding, after which it left the judgment to the surgeon; the researchers said that this phenomenon was not abnormal bleeding. The surgery was not completed entirely independently by the AI. A licensed veterinarian was responsible for the incision, craniotomy, and suturing and supervised the entire process. Irreversible critical steps required human approval, and the equipment was also equipped with motion range limits, injection dose limits, and an emergency stop switch.
At present, the relevant experiments are still in an early exploratory stage. The research lead acknowledged that the sample size of three operations is limited and is not yet sufficient to prove that the AI can reliably master the relevant skills; whether the treatment plans chosen by different models are effective also needs to be determined after observing mouse behavior and brain activity. These preliminary experiments demonstrated the possibility of AI moving from analyzing scientific research data and formulating experimental plans to directly operating real experimental equipment, but the results have not yet been fully independently verified or peer-reviewed, and there is still a long way to go before application in human brain surgery.