Neuralink cuts some implant users' calibration from 10 minutes a day to 10 minutes a week

By Inside BCI · October 3, 2026

Participants in Neuralink’s trial typically began each day with about 10 minutes of calibration, a routine that lets the software relearn how their brain signals map to cursor movement before they can use their computer. Across participants it averaged 55 minutes a week. Some now calibrate for 10 minutes a week. Neuralink published the change on 1 October.

What Neuralink changed

Since its clinical trial began, participants have streamed more than 50,000 hours of brain activity recorded during everyday use of the implant, none of it labelled with what the person was trying to do at the time. The first participant, Noland Arbaugh, accounts for more than 9,000 hours on his own. Until now almost none of it was used. Each participant’s decoder, the software that turns brain signals into intended actions, was trained on short labelled sessions recorded during calibration.

Neuralink has now pretrained participant-specific models, each on thousands of hours of one person’s own recordings. The model learns the recurring structure in their brain activity without labels and passes the decoder a steadier signal, one that drifts less from day to day.

What changed for participants

Participants who recalibrate at the first sign of slipping control went more than a week without wanting to. Some decoders held strong performance for over three weeks, and in one test a participant had usable control from a calibration done 20 months earlier. On a simulated robotic arm, where decoders could previously become unusable within six days, decoders stayed controllable a week later.

Six participants set personal bests on Neuralink’s cursor test, which scores speed and precision in bits per second. Three beat the company’s previous record of 10.39, and one participant, P15, reached 11.32, which Neuralink calls a record for a brain-computer interface. The median across participants is about 10. Labs score cursor control on different tasks, so rates compare loosely across groups. Paradromics reported information rates above 200 bits per second in October 2025 from its Connexus implant in two sheep, measured offline by decoding, from auditory-cortex activity, which tones were played to the animals.

Why the hours matter

The hours accumulate as a byproduct of everyday use, so the pool grows with every participant and every day they use the device. A rival with fewer implants in daily use has fewer hours to train on. Calibration time is a running cost borne by the user of any implant sold as a home device, and moving it from close to an hour a week to 10 minutes a week, as some participants now do, reduces what the product asks of the people who live with it.

The advantage of scale across people is still unproven. Every live result in the update came from a model trained on one participant’s data, and models trained on several participants together have so far performed no better in live use than single-participant models. In one experiment, a model trained only on participant P9’s data, with 99 percent of its parameters held fixed, outperformed P2’s own existing decoder.

The platform Neuralink is describing

Beyond pooling data across participants, Neuralink set out three further goals. The first is calibrating once and keeping control for a year or more. The second is never calibrating, so that a person can, in the company’s words, “wake up from surgery with a device that matches your intent.” The third is a shared decoder for all intended hand movement, which Neuralink calls an API for the motor cortex: users would calibrate once across applications, and developers would build on a common interface, which the company says would pave the way for a BCI app ecosystem. If opened to outside developers, a shared decoder would make Neuralink’s interface the layer that applications build on. Neuralink is hiring machine-learning engineers for the work.

Synchron is making the same bet

Synchron unveiled a roadmap to Chiral in March 2025, alongside Nvidia’s GTC conference, a brain foundation model to be trained with Nvidia on de-identified neural data. Chief executive Tom Oxley tied the plan to scale: “This is possible because of our ability to scale large datasets, by making BCI as common as a stent insertion.” Synchron cited 20 patient-years of implant experience since 2019 at the time, a measure of time implanted.

Not disclosed

How many participants contributed to the 50,000 hours, and how many now calibrate weekly, is not disclosed. The consent terms covering the use of participants’ everyday recordings for model training, and whether the data are de-identified, are not public. The results are reported by Neuralink and have not been independently verified.

What to watch

Whether models pooled across participants beat single-participant decoders in live use, the test of whether Neuralink’s data volume compounds.

Whether the API for the motor cortex becomes a programme open to outside developers.

Whether Neuralink submits the results to a peer-reviewed journal or a shared benchmark.

Whether Synchron reports live user results from Chiral.

Sources

Primary:

Coverage:

Cross-reference to prior InsideBCI coverage:

news Neuralink PRIME-trial foundation-model neural-decoder calibration bits-per-second neural-data Noland-Arbaugh API-for-the-motor-cortex Synchron Chiral Nvidia Paradromics SONIC-benchmark