Summary

• In a new scientific preprint, Paradromics introduces SONIC, a rigorous, open benchmarking standard to measure the performance of any brain-computer interface (BCI).
• Using this standard, the Paradromics Connexus® BCI achieves an information transfer rate over 200 bits per second (bps) with negligible delay. This is over 20 times faster than the initial reported performance of other intracortical systems like Neuralink’s and orders of magnitude beyond endovascular systems like Synchron’s.
• We encourage other BCI companies to benchmark their devices using SONIC or similarly robust methods. By fostering a transparent method for comparing performance, we pave the way for innovation in human-computer interaction.

Why a new standard will accelerate the entire BCI industry

Brain-computer interfaces are complex hardware-software stacks that take millions of dollars and many years to bring to market. As platform technologies, a single device can enable multiple applications. The challenge is that the highest performing BCIs are surgically implanted medical devices and testing them involves costly and slow clinical trials for specific applications.

A close example with similar challenges comes from the semiconductor industry, where new chip commercialization requires significant time and capital. To drive progress, that field adopted benchmark tests, engineering assessments that reflect underlying system properties. These tests, while not a substitute for final user testing, allow for faster, more objective, and more comprehensive feedback loops during design and give confidence to the end-user when selecting a chip. The BCI industry has lacked such a standard.

With this in mind, Paradromics developed the Standard for Optimizing Neural Interface Capacity (SONIC) benchmark. Engineering tests are essential in advance of clinical tests to ensure the BCI system will meet performance requirements. Application-agnostic performance metrics are critical for moving the field forward, especially when they can be demonstrated preclinically.

Why Achieved information transfer rate and latency both matter

Theoretical frameworks for information transfer rely on assumptions that often don't hold true for complex neural systems. The most honest method is to report the data you actually transferred.

But a high information transfer rate alone isn’t enough. It's possible to game the results without accounting for delay (latency). Some decoding methods record long blocks of data and look backward in time, reporting a subset of application-specific metrics that are impressive at a surface level, but are achieved with compromises, including the introduction of long delays. Many applications, like conversational speech, can't tolerate such delays.

Accounting for delay helps explain discrepancies between BCI trials that report similar "words per minute" despite using devices with vastly different information transfer rates. In pre-recorded demos, it's easy to hide factors like latency and performance variability in post-production.

The new record: Setting the bar for the industry

Our benchmark testing of the fully implantable, wireless Connexus BCI pushed the limits on information collected per second while measuring error rates and delays. We made these measurements over 10 months after device implantation. The current scores to beat are:

• 200+ bits per second with 56ms total system latency‍
• 100+ bits per second with 11ms total system latency

These rates are 10-20 times higher than demonstrated in BrainGate or Neuralink clinical trials and 100-200 times faster than Synchron's reported performance. Importantly, they exceed the information transfer rate of transcribed human speech (~40 bps), giving high confidence in our ability to deliver a high-performing communication BCI in clinical trials starting next year.

What high performance feels like: The demo

Numbers don't always convey the real-world impact. To make the difference tangible, we created a few demonstrations.

1. Visualizing 200 bps Performance

In our animation, the Connexus BCI decodes fluid text at the reading rate of a skilled reader. For this demonstration, we played sound sequences (5 tones mapped to 1 letter) to the sheep, whose neural activity was collected with the Connexus BCI, decoded, and transformed back into characters. For comparison, on the right, we show the same task at the representative rate of BCI outputs through the Neuralink device (e.g. an alphabet WebGrid task) or the approximate speed of academic intracortical studies using the Utah array. In the final example on the bottom, we display the text at the representative rate of BCI outputs through Synchron’s device based on reported outcomes. The difference in user experience is immediately apparent.¹

2. Why delay matters: Super Mario Bros.^TM Wonder

The above text animation shows the impact of data throughput, but delay is equally critical. Our Super Mario Bros. Wonder demo shows this intuitively. With zero added delay, gameplay is fluid. At 200ms, it becomes clumsy. At a delay of 500ms, the game is unplayable. Applying SONIC benchmarks, the Connexus BCI achieved over 100 bps with a negligible latency of 11ms.

3. Experience it yourself: The sheep run game

Our interactive Sheep Run game lets you directly experience how delay and error rates affect control. Adjust the sliders and see how performance changes. It makes the abstract tangible and highlights why maximizing information throughput is critical for the user. The level of information required to control this game is low, and trivial to achieve for intracortical BCIs like the Paradromics Connexus, Neuralink device and the Utah Array. However, for devices that have lower information transfer rates (such as intravascular devices), control may require compromises. Specifically, for such devices the user would have to tolerate either high delay or high error. You can simulate these conditions using the provided preset buttons.

Note: You can increase the impact of delay by decreasing visibility, hiding the upcoming cactuses and birds in “fog.” While a user can develop compensation strategies for delay, with low visibility - future obstacles cannot be easily anticipated and compensation strategies fail.

How the SONIC Benchmark works

We conduct our preclinical experiments in sheep. For the SONIC benchmark, we play controlled sequences of sounds and use the fully implanted Connexus BCI to predict which sounds were presented based on the recorded neural activity from the sheep's auditory cortex. By calculating the mutual information between the sounds presented and the sounds predicted, we get a true measure of information transfer rate. A full definition of the SONIC benchmark is available in the scientific preprint.

Next steps: The pressing need for an industry standard

As the field of BCI has matured, the community is actively developing clinical outcome assessments for BCI devices. This industry-wide coordination has led some to wonder whether device designers should move away from engineering-focused metrics toward more traditional clinical endpoints. We assert that both engineering and clinical tests are important to build robust platforms. As end-to-end systems are brought to market for specific applications, BCIs must be designed to meet specific user needs. But as these complex BCI platforms take years to develop and have the potential to be used for many indications, it is also clear that developing application-agnostic performance metrics will move the entire field forward, especially if these tests can be implemented preclinically.

This is a call to action. We advocate the adoption of SONIC’s robust benchmarks or similarly rigorous methods because the entire field advances faster when guided by transparent, standardized performance metrics. Publicly reported benchmarks will accelerate innovation, can lead to better outcomes for BCI users, and unlock a future of possibilities.

The Paradromics Connexus BCI has set a new industry standard in performance. But just as important, we have established a rigorous framework for measuring that performance. With record-setting performance, we have confidence in Connexus BCI’s ability to drive advanced applications and serve as the foundation for an industry-leading platform for human-technology integration.

Read the full scientific preprint: Perkins, S.M., et al. (2025). SONIC: A Benchmarking Paradigm for Brain-Computer Interfaces. bioRxiv.

For media inquiries, please contact: media@paradromics.com

_______

¹ How it works: Decoding Speech Transcripts from Brain Signals 

We encoded the transcribed speech as characters. Each character was assigned a unique five-note musical tone sequence to make a “dictionary”. These characters were then transmitted one after another in a continuous stream.

We then decoded the text from neural signals. We first identified individual tones, and matched each five-tone sequence to the most similar character in the dictionary. We then calculated the transmission speed by measuring the amount of useful information sent per second.

The Trade-Off Between Speed and Accuracy 

The final transmission speed for characters was very high, over 5x faster than other similar systems. While still fast, this rate is slower than the maximum capacity of Connexus for transmitting the individual tones. This difference in speed is because we intentionally used longer, five-tone sequences for each character to achieve near-perfect accuracy.

This highlights a classic engineering trade-off. For instance, if we used shorter tone sequences, we could have significantly increased the speed, but at the cost of reduced accuracy. Because Connexus has such a high information transfer capacity, we had the flexibility to prioritize accuracy over raw speed while still maintaining excellent performance. Further, character-based encoding was used due to its simplicity and direct connection to current onscreen typing demonstrations by Neuralink and Synchron, although other encoding schemes could have simultaneously enhanced both speed and accuracy.