Neuralink Sets New BCI Record After Training on 50,000 Hours of Brain Data

Neuralink says it has begun putting more than 50,000 hours of neural recordings from clinical trial participants to work, helping improve cursor control, reduce calibration time and set a new brain-computer interface performance record.

In an October 1 blog post, Neuralink said participants have streamed more than 50,000 hours of unlabeled, freeform neural data during the two years since its clinical trial began. Its first participant alone recorded more than 9,000 hours, representing 22.4 billion neural spikes.

Until recently, Neuralink says nearly all of that information was going unused. The company has now trained participant-specific neural encoders on thousands of hours of recordings collected from each person before using those models for cursor control.

The approach lets the system learn recurring patterns in a participant’s brain activity from past recordings instead of starting each task from scratch. Neuralink says the encoders turn noisy neural signals, which can shift over time, into more stable representations called embeddings that can make decoding more accurate and durable.

That has cut the amount of calibration some participants need. Users previously spent an average of 55 minutes per week recalibrating their systems, typically about 10 minutes per day. Some participants have now reduced that to about 10 minutes per week.

One participant used the same decoder for five consecutive days while reaching 10 bits per second, or BPS. Some decoders remained strong for more than three weeks, while one participant was still controlling a decoder more than 1.5 years later.

Six participants also set personal BPS records. Neuralink says the median across participants is roughly 10 BPS, while three exceeded the previous 10.39 BPS record. One participant reached 11.32 BPS, which the company described as a new BCI record.

P9 said the new model was “Better than the prior model. Smoother, more directionally accurate, felt effortless. Didn’t have to push or anything, it just glided where I wanted it to go.”

Before reaching 11 BPS, P15 said, “[Control] just feels better. [Click behavior] is better for games, quick and responsive. Drop sensitivity a tiny bit and I could easily hit 11.”

Neuralink also reported improvements to click decoding, including fewer accidental clicks and less reliance on software corrections such as smoothing and low-velocity suppression. An experimental “cursor teleportation” system, which predicts an intended target location directly, crossed 10 BPS during testing.

The latest results build on Neuralink’s expanding clinical work. Earlier this year, 21 participants were enrolled as Neuralnauts, following progress detailed in the company’s Summer 2025 update.

Neuralink’s longer-term goal is to combine data from multiple participants into a shared, drift-resistant neural foundation model. In one experiment, the company transferred a model from P9 to P2 while keeping 99% of its weights frozen, and that model still outperformed P2’s existing decoder.

Next, Neuralink says it is exploring one-shot and zero-shot calibration, along with a shared “API for the motor cortex.” The broader aim is to move beyond models trained separately for each participant toward a system that can transfer more of what it has learned across users while remaining stable over time.