Sabi Cap is the name of a baseball cap that Palo Alto startup Sabi says can read your brain activity and turn the words you imagine into text, with no implant and no gel. On 9 October 2026 the company announced a $50 million seed round led by Khosla Ventures, and it made a striking new claim: its model can already predict a person’s next three to four keystrokes from brain signals before they are typed.
The Sabi Cap keystroke claim came via TestingCatalog’s explainer and its post on X, which is where the headline above comes from. It is the most concrete performance statement Sabi has made since it came out of stealth in April. It is also, so far, unsupported by any published accuracy figure or independent test.
This article sets out what Sabi announced, how the Sabi Cap is supposed to work, what “predicting keystrokes” does and does not prove, how the Sabi Cap compares with brain-to-text systems that have published results, how the plan has changed since April, and why neural data privacy is the question every buyer and employer should ask first.
Table of contents
- What Sabi Announced About the Sabi Cap
- How the Sabi Cap Reads the Brain Without Touching the Scalp
- What Predicting Keystrokes Really Means for the Sabi Cap
- How the Sabi Cap Compares With Other Brain-to-Text Systems
- From Beanie to Baseball Cap: How the Sabi Cap Plan Changed
- Neural Data Privacy: The Sabi Cap’s Hardest Problem
- What the Sabi Cap Could Mean for Work and Accessibility
- What We Still Don’t Know About the Sabi Cap
- Sabi Cap Questions Answered
- References
What Sabi Announced About the Sabi Cap
The announcement is mainly a funding story with a product roadmap attached. The press release, distributed through GlobeNewswire and carried on Yahoo Finance, describes the Sabi Cap as a “brain-AI interface designed for everyday wear” that “looks like an ordinary baseball cap”.
The $50 million seed round
Khosla Ventures led the round. Accel, Initialized Capital, former OpenAI chief product officer Kevin Weil, DST Global, Collaborative Fund and Ascend also took part. Sabi says it will spend the money on expanding its neural dataset, finishing its custom chip, manufacturing early devices and hiring in neuroscience, hardware and machine learning. Chief executive Rahul Chhabra confirmed the raise on X the same day.
Vinod Khosla supplied the line that frames the whole pitch: “You cannot perform a billion brain surgeries to replace a billion keyboards. Every computing era is defined by its interface.”
Three ways to talk to AI without typing
The release lists three forms of interaction the Sabi Cap is being built for:
- Imagined speech to text: words a user deliberately imagines appear on a connected screen without being spoken or typed.
- Thought to prompt: a trained mental command becomes an instruction to an AI assistant, such as drafting a message or retrieving information.
- Intent prediction: the system anticipates likely keystrokes or on-screen actions and presents them “for the user to complete or confirm”.
The Sabi Cap connects to a phone or computer wirelessly, “much like a pair of earbuds”. Sabi plans to show early versions of these interactions at CES 2027 in Las Vegas.
Who is building it
Sabi was co-founded by Chhabra and chief technology officer Atmadeep Banerjee. Banerjee is a co-first author of MindEye, a NeurIPS 2023 paper that reconstructed images people were viewing from fMRI brain scans. The company says its team includes hardware, machine learning and design leaders from Kernel, Apple, Microsoft, Meta, Adobe and Nike, alongside neuroscientists from Stanford.
| Item | What Sabi says | Source |
|---|---|---|
| Funding | $50 million seed, led by Khosla Ventures | Press release, 9 Oct 2026 |
| Form factor | Baseball cap with sensors hidden inside | Press release |
| Sensor size | 1 to 5 millimetres, reading through hair across a gap of up to 5 mm | Press release |
| Chip efficiency | More than 50 times as power-efficient as TI’s ADS1299, in internal benchmarks | Press release |
| Dataset size | More than 100,000 hours of labelled, non-invasive recordings | Press release |
| Keystroke prediction | Next three to four keystrokes, before they are typed | TestingCatalog |
| Target speed | About 30 words per minute at first | WIRED, April 2026 |
| Next milestone | Early prototype demonstrated at CES 2027 | Press release |
How the Sabi Cap Reads the Brain Without Touching the Scalp
The Sabi Cap relies on electroencephalography, or EEG, which measures the tiny voltage changes produced when large groups of neurons fire. EEG has been used in hospitals and labs for a century. Its weakness is the distance between the sensor and the source.
Why EEG is hard to decode
As TestingCatalog explains, brain signals reaching the scalp “arrive weak and blurred” after passing through bone, tissue and hair, and muscle movement and electrical noise “can drown them out”. That is why implanted electrodes, which sit next to neurons, have produced the most accurate brain-to-text results so far. Conventional EEG also needs gel and electrodes pressed against the scalp, which rules out casual daily use.
A non-contact EEG chip
Sabi’s answer is custom hardware. It designs its own sensors and application-specific chips, and the centrepiece is what it calls the first non-contact EEG chip, designed in-house and fabricated by TSMC. The release says each Sabi Cap carries sensors measuring one to five millimetres, designed to capture signals through hair “across a gap of up to five millimeters”.
The comparison Sabi chose is telling. It says its chip was “more than 50 times as power-efficient” in internal benchmarks than the Texas Instruments ADS1299, the eight-channel, 24-bit converter that sits inside many research-grade EEG systems. Power matters because every milliwatt becomes heat and electrical noise against the wearer’s head.
Why the sensor count forces a custom chip
At its April launch Sabi told WIRED the Sabi Cap would carry “anywhere from 70,000 to 100,000 miniature sensors”, against a dozen to a few hundred in most EEG devices. Here is a simple way to see why off-the-shelf parts cannot do that. One ADS1299 handles eight channels, so reading 70,000 to 100,000 sensors with that chip would take 8,750 to 12,500 of them (70,000 divided by 8, and 100,000 divided by 8). No one could fit that into a hat. Whatever the final channel count, a custom chip is a necessity, not a nicety.
| Sensor count | ADS1299 chips needed (8 channels each) | Typical setting |
|---|---|---|
| 8 | 1 | Consumer headband or simple research rig |
| 256 | 32 | High-density research EEG cap |
| 70,000 | 8,750 | Low end of Sabi’s April figure |
| 100,000 | 12,500 | High end of Sabi’s April figure |
The October release does not repeat the 70,000 to 100,000 figure. It describes sensor size and the chip, but not how many sensors a finished Sabi Cap will carry. That is worth asking about before CES.
The Brain Foundation Model
The software half is what Sabi calls its Brain Foundation Model. The release says it was trained on “more than 100,000 hours of labeled, non-invasive recordings collected while participants imagined words and performed specific mental tasks”, which Sabi calls “the largest disclosed dataset of its kind”. In April, Chhabra told WIRED the company had collected 100,000 hours from 100 volunteers, which works out at about 1,000 hours per person.
The goal is a model that learns patterns shared across many people, so a new wearer can use the Sabi Cap without long calibration sessions. In machine learning terms this is a transfer learning problem: knowledge learned from one group of brains has to carry over to a brain the model has never seen. JoJo Platt, a neurotechnology consultant quoted by WIRED, put the consumer requirement simply: “These devices are going to have to be ready to go out of the box.”
What Predicting Keystrokes Really Means for the Sabi Cap
The headline claim is that Sabi’s model “can already predict a person’s next three to four keystrokes from brain signals before they are typed”. It sounds like mind reading. It needs careful reading.
Prediction is easier than decoding
Predicting the next few keys of a sentence is something your phone’s keyboard already does, using nothing but the text typed so far. A language model that sees “Thanks for your” can guess the next characters with high confidence without any brain data. So the meaningful question is not whether the Sabi Cap can predict keystrokes, but how much better it predicts them with neural signals than a text-only model does without them.
Sabi has not published that comparison, nor an accuracy rate, nor the conditions of the test. The release itself is careful: intent prediction will “anticipate likely keystrokes or on-screen actions and present them for the user to complete or confirm”. That describes an assistive autocomplete, with the user in control, rather than a decoder that types for you.
What a fair test would report
If Sabi wants the keystroke claim to carry weight, it should publish four numbers:
- Accuracy per predicted key, for the first, second, third and fourth keystroke ahead.
- A text-only baseline using the same language model without brain signals, so the neural gain is visible.
- Results on new wearers the model never trained on, which is the whole point of a foundation model.
- Latency, because a suggestion that arrives after you have typed the key is useless.
Why the claim still matters
None of this means the claim is empty. Even a small, reliable neural signal about upcoming keys could speed up typing for people with motor impairments, where every saved keystroke counts. And intent prediction is a far easier first product than open-ended thought-to-text. It is a sensible place for the Sabi Cap to start. It is just not yet evidence that the Sabi Cap reads thoughts.
How the Sabi Cap Compares With Other Brain-to-Text Systems
Brain-to-text research has moved quickly, and several systems have published numbers. Comparing them shows the size of the gap Sabi is trying to close.
Meta’s Brain2Qwerty: the closest published benchmark
Meta’s research lab, working with Spain’s Basque Center on Cognition, Brain and Language, published Brain2Qwerty in February 2025. It decoded sentences from the brain activity of 35 healthy volunteers while they typed. With magnetoencephalography (MEG), it reached an average character error rate of 32%, and 19% for the best participants. With EEG, the technology the Sabi Cap uses, the error rate was 67%.
The second version, reported by The Next Web in July 2026, reached average word accuracy of 61% with MEG, rising to 78% for the best participant, after training on about 22,000 sentences from nine volunteers who each wore the scanner for about 10 hours. Its EEG character error rate was still 65%. The system is not real-time, needs a room-sized scanner and learns from people who are physically typing.
Implants are still far ahead
At Stanford, a 2023 Nature study decoded the attempted speech of a woman with ALS through electrode arrays implanted in her brain. She reached 62 words per minute, with a 9.1% word error rate on a 50-word vocabulary and 23.8% on a 125,000-word vocabulary. The paper puts natural conversation at about 160 words per minute, and earlier hand-movement BCIs at 8 to 18 words per minute.
| System | Signal | Surgery? | Published result |
|---|---|---|---|
| Sabi Cap | Non-contact EEG | No | None yet; targets about 30 words per minute |
| Brain2Qwerty v1 (Meta, 2025) | EEG or MEG while typing | No | Character error 67% on EEG, 32% on MEG |
| Brain2Qwerty v2 (Meta, 2026) | MEG while typing | No | Word accuracy 61% average, 78% best; EEG character error 65% |
| Stanford speech BCI (Nature, 2023) | Implanted microelectrode arrays | Yes | 62 words per minute; 23.8% word error, large vocabulary |
| Meta Neural Band (2025) | Wrist muscle signals (EMG), not brain | No | Shipping as a control band for Meta Ray-Ban Display glasses |
The last row is a useful reminder that non-invasive input is already on sale. Meta’s Neural Band, launched in September 2025, reads the electrical signals of muscles in the wrist, which are far stronger than brain signals at the scalp. The Sabi Cap is aiming at something much harder.
Bar widths are each speed divided by 160. Sabi’s 30 words per minute target would be about half the Stanford implant’s speed (30 divided by 62 is 0.48) without any surgery. That would be a remarkable result for EEG, which is exactly why it needs independent evidence.
The data gap is real, and it favours Sabi
The one area where Sabi plainly leads is data volume. Brain2Qwerty v2 trained on roughly 90 hours of recordings (nine volunteers at about 10 hours each). Sabi claims more than 100,000 hours, over a thousand times as much. Meta’s own result supports the idea that more data helps: The Next Web reports Meta’s view that accuracy “climbs steadily” as more data is added. Whether scale can overcome EEG’s physics is the bet Khosla and the other investors are making.
From Beanie to Baseball Cap: How the Sabi Cap Plan Changed
Sabi’s April debut and its October announcement describe the same company, but not quite the same plan. Comparing them is the clearest way to judge the timeline.
| Topic | April 2026 (WIRED, New Atlas) | October 2026 (press release) |
|---|---|---|
| First product | A beanie, with a baseball cap version to follow | A baseball cap |
| Availability | Beanie “available by the end of the year” | Early prototype shown at CES 2027; units prepared for the waitlist |
| Sensors | 70,000 to 100,000 miniature sensors | Sensors of 1 to 5 mm; count not stated |
| Dataset | 100,000 hours from 100 volunteers | More than 100,000 hours, labelled |
| Funding disclosed | Investors named, amount not given | $50 million seed |
| Performance claim | About 30 words per minute at first | Predicts next three to four keystrokes (TestingCatalog) |
The schedule moved
In April, Chhabra told WIRED that the beanie would be available by the end of 2026, and New Atlas repeated that it was “slated to arrive by the end of 2026”. In October, the first product is the Sabi Cap, and the next milestone is an early prototype at CES 2027, which opens in January. Hardware schedules slip all the time, and a pivot to a more wearable form factor is a reasonable choice. But anyone planning around a 2026 delivery should now assume 2027 at the earliest.
The claims became more cautious
The October release is notably measured. It speaks of words a user “intentionally imagines”, of the Sabi Cap being used “when they intentionally want to communicate with AI”, and of keystrokes presented “for the user to complete or confirm”. That language reads like a company preparing for regulators and privacy questions as well as investors.
Neural Data Privacy: The Sabi Cap's Hardest Problem
Chhabra put it plainly to WIRED in April: “neural data is the most private kind of data that a person could possibly have.” A device that records brain activity all day raises questions no keyboard ever did.
What Sabi says it will do
According to the release, neural data is encrypted “beginning at the sensor”, and Sabi’s system “is designed to run inference while that data remains encrypted, preventing servers from receiving a user’s unencrypted neural signals”. The company is also building controls over when the Sabi Cap records, which applications may receive commands and which actions need confirmation. In April it said it was consulting neurosecurity experts from Stanford and elsewhere to audit its technology stack.
Running AI inference on encrypted data is possible but computationally expensive, and Sabi has not said which technique it uses. That detail matters for cybersecurity reviews, because “encrypted in transit” and “never decrypted on the server” are very different promises.
What the law says
Neural data is already regulated in parts of the United States. A Davis Wright Tremaine analysis notes that California, Montana, Colorado and Connecticut have amended their privacy laws to cover it, with different definitions and obligations. Federally, Senators Cantwell, Schumer and Markey proposed the MIND Act in September 2025, which would direct the Federal Trade Commission to study neural data and recommend rules. Internationally, UNESCO’s Recommendation on the Ethics of Neurotechnology was adopted in November 2025.
| State | What counts as neural data | Consent model |
|---|---|---|
| Colorado | Central and peripheral nervous system, including derived data | Opt-in, when used to identify a person |
| Connecticut | Central nervous system only | Opt-in for sensitive data |
| California | Central and peripheral, excluding algorithmically derived data | Right to opt out of certain processing |
| Montana | “Neurotechnology data”, excluding downstream physical effects | Not summarised in the comparison |
The table summarises the Davis Wright Tremaine comparison. For UK and European organisations, brain activity used to infer health or identity would almost certainly be special category data under GDPR, which brings stricter conditions and a data protection impact assessment.
The questions to ask before anyone wears one at work
- Is raw neural data stored, and if so, where, for how long and under which jurisdiction?
- Is user data used to train the Brain Foundation Model, and can a user opt out?
- What exactly can an application connected to the Sabi Cap receive: text, commands or signals?
- Can an employer see or receive anything the wearer did not choose to send?
What the Sabi Cap Could Mean for Work and Accessibility
If it works, a cap that turns intended words into text would matter most to people who cannot easily type or speak. For everyone else, the use cases are more speculative but still interesting.
Accessibility comes first
Brain2Qwerty’s main limitation, as The Next Web noted, is that it learns from people who are typing, while its intended users often cannot type at all. Sabi trains on imagined words and mental tasks instead, which, if it works, would suit people with motor impairments far better. That is the most valuable test case for the Sabi Cap and the one where independent clinical evaluation would be most persuasive.
Silent input for AI agents
Sabi’s pitch is about AI more than accessibility: “thought to prompt” for assistants and agents. Speaking to an assistant in an open-plan office is awkward, and typing is slow on a phone. We have covered the same push towards ambient input in Google’s offline AI dictation app and in Apple Watch’s always-listening features. A Sabi Cap that does the same silently would be the logical next step, and a much bigger privacy question.
What businesses should do now
There is nothing to buy yet, so the practical steps are about policy:
- Add neural and biometric wearables to your acceptable-use and bring-your-own-device policies before staff start turning up with them.
- Treat any workplace pilot as high-risk processing and run an impact assessment first; our data protection guidance covers the basics.
- Ask vendors the questions in the privacy section above, in writing.
- Watch for independent results after CES 2027 rather than relying on launch videos.
What We Still Don't Know About the Sabi Cap
Funding announcements are written to be optimistic. Here is what the October release and the coverage leave open.
| Question | Status |
|---|---|
| Accuracy of text decoding and of keystroke prediction | Not published |
| Words per minute actually achieved | Target of about 30 stated in April; no result |
| Peer-reviewed or independent testing | None yet, as TestingCatalog notes |
| Price and shipping date | Not announced |
| Number of sensors in the finished cap | Not restated since April |
| How encrypted inference works | Technique not disclosed |
| Battery life and comfort over a working day | Not disclosed |
The questions CES should answer
The CES 2027 demonstration is the first real test of the Sabi Cap. Look for live typing by people who are not Sabi employees, a side-by-side comparison with ordinary autocomplete, and a clear statement of how long each wearer trained the system before the demo.
The wider brain-and-AI story
Sabi sits within a broader wave of research connecting brains and AI models, from AI models built from rat brain neurons to the argument that cognitive science should shape how AI is built. The Sabi Cap is the most consumer-facing bet of the group, which is why the evidence bar should be high.
Sabi Cap Questions Answered
Can the Sabi Cap read my thoughts?
Not in the sense of reading anything you think. Sabi says it decodes words a user “intentionally imagines” and acts when they “intentionally want to communicate with AI”. It has not published evidence of how well that works.
How fast can you type with the Sabi Cap?
Sabi’s April target was about 30 words per minute, improving with use. No measured speed has been published. For comparison, an implanted speech BCI reached 62 words per minute in a 2023 Stanford study.
Does the Sabi Cap need surgery or gel?
No. It uses non-contact EEG sensors inside a cap and reads through hair across a small gap, according to the company.
When can I buy a Sabi Cap?
There is no price or shipping date. Sabi plans to show an early prototype at CES 2027 and is preparing units for people on its waitlist.
How is the Sabi Cap different from Neuralink?
Neuralink implants electrodes in the brain through surgery, which gives much stronger signals. The Sabi Cap is worn like a hat and reads weaker signals through the skull, trading accuracy for convenience and safety.
References
Sabi raises $50 million seed round led by Khosla Ventures (GlobeNewswire via Yahoo Finance)
How Sabi’s brain-reading cap turns thoughts into text (TestingCatalog)
This beanie is designed to read your thoughts (WIRED)
This mind-reading beanie could make keyboards obsolete (New Atlas)
Brain-to-text decoding: a non-invasive approach via typing (arXiv)
Meta’s AI reads typed sentences from the brain, no surgery required (The Next Web)
A high-performance speech neuroprosthesis (Nature)
Reconstructing the mind’s eye: fMRI-to-image with contrastive learning and diffusion priors (arXiv)
ADS1299 low-noise 8-channel 24-bit ADC for biopotential measurements (Texas Instruments)
Meta Ray-Ban Display: AI glasses with an EMG wristband (Meta)
U.S. senators propose MIND Act to study neural data standards (Davis Wright Tremaine)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
