Grok Voice Transcribe 2.0 was released on 18 September, priced at $0.10 per hour of audio for batch work and $0.20 per hour for streaming. Both figures are unchanged from the previous version.

The announcement is published under SpaceXAI, the name xAI has carried since July after SpaceX acquired it in February. Coverage still calls the company xAI, and its own footer does not.

Read the accuracy claim precisely

The company says the model is twice as accurate as Grok Voice Transcribe 1.0. That is a comparison with its own predecessor, not with anyone else’s model.

Against the field it makes a narrower claim, that it ranks first for accuracy among 32 streaming models on the public Artificial Analysis leaderboard, and describes itself as one of the most accurate rather than the most accurate. The hedge is the company’s own and is worth keeping.

Its internal charts compare against ElevenLabs Scribe v2 and Deepgram Nova-3. Those are vendor-run evaluations against vendor-chosen comparators, which is normal and is not the same as an independent test.

The price is the actual news

Holding price while claiming a doubling of accuracy is a statement about the market rather than the model. Speaker labelling, word-level timestamps and key term biasing are all included at no extra cost.

Those were billable features not long ago. Transcription is being priced as a commodity input, and the companies selling it are competing on what they can bundle.

The squeeze is general. OpenAI has pushed out new voice API models, and Microsoft’s in-house transcription model was benchmarked against Whisper, Gemini and ElevenLabs across 25 languages.

Where the advantage actually comes from

SpaceXAI is explicit about it. The model is trained on what it calls a unique dataset of live, noisy, multilingual audio recorded across a diverse set of environments.

It also describes the pipeline that produces that audio. Grok Voice powers tens of thousands of customer-support calls a day, transcribes millions of hours of video narration, and runs the assistant in Tesla vehicles.

Clean-speech accuracy is solved. What is left is telephony, accents, crosstalk and in-car commands, and the edge there belongs to whoever has the most real conversations flowing through their systems.

The evaluation sets are the detail to notice

The company says it measures word error rate on four internal sets drawn from production traffic. They are customer-support telephony, conversations with Grok, short multilingual voice commands, and spoken credentials.

That last set is described as account codes, phone numbers, email addresses and addresses read aloud. It is a standing evaluation corpus of people reading out their own identifying details.

Nothing in the announcement says how that material is obtained, retained or consented to. It may well be covered by enterprise terms, and the announcement does not say.

The European question that follows

A customer-support call has two parties and only one of them has a contract with the AI vendor. The caller reading out an account number has agreed to nothing.

Voice recordings are personal data in Europe, and a set built specifically around spoken credentials is about as sensitive as transcription data gets. Purpose limitation and retention are the obvious questions, and they are answerable.

The practice is not new, which is part of the point. People have been listening to supposedly automated transcripts for over a decade, and the disclosures have consistently lagged the practice.

Where the real gain is

The headline improvement is multilingual, and specifically short phrases. On the company’s own 19-language set of voice-assistant utterances, word error rate falls from 20.6% to 6.8%.

Short commands give a model almost no context to identify the language, which is exactly the in-car problem. The largest reported gain lands on the use case SpaceXAI owns outright through Tesla.

Multilingual is also where European vendors have concentrated. DeepL runs real-time voice translation across more than 40 languages, and ElevenLabs has listed transcription among five services on the UK government’s cloud framework.

The customer proof point

Atlassian says it found the model more accurate than its existing supplier and now transcribes every Loom video with it. That is a real switch by a named buyer, which is worth more than a benchmark chart.

Version 1.0 is being deprecated within weeks, with pinning available during the transition. Customers on the old model are being moved whether or not they asked.

What to watch

Watch whether anyone asks about the credentials corpus. It is the one genuinely novel disclosure in the announcement and nobody has picked it up.

Watch the price floor. At 10 cents an hour the model is nearly free, and the business is the audio flowing through it.

Last night, in a clip viewed more than eight million times, AI actress Tilly Norwood glitched mid-interview and unexpectedly began speaking Chinese on the Piers Morgan Uncensored show.

Norwood has been the subject of much controversy and mainstream commentary for starring in an upcoming AI-generated film, Misaligned.

Her creators run a “Talking Tilly” service that lets anyone video-call the AI character behind the viral moment, so today I tried it for myself, and read the fine print most callers won’t.

A face scan before you can say hello

Before your first call connects, you must pass an automated age check.

A video selfie is analysed by Didit, a Spain-based identity verification provider, to estimate your age, with a government photo ID upload as the fallback if the estimate is unclear.

Xicoia Ltd, the UK company behind Tilly, states the selfie travels from your device directly to Didit, that no faceprint or biometric template is created, and that neither the selfie nor any ID image is kept after the check; the company says it retains an approximate age band and a reference number instead.

The check cannot be skipped, applies to callers worldwide, and was only added to the service’s terms this month, alongside a workplace-use ban and new automated safety systems.

The face scan is not the only analysis running.

During every call, the system watches your camera feed and listens to your tone of voice to infer your emotional state, so the character can, in the company’s words, respond in a way that fits the mood.

The privacy policy is candid that this “cannot be switched off for an individual call.” If you don’t want it, don’t call.

Notably, both the age check and the mood-sensing rely on legitimate interests rather than consent as their legal basis, a choice Xicoia’s own version history shows was made in September.

Calls are recorded, transcribed, and processed live by US providers, with the character’s responses generated by Google’s Gemini model via conversational video platform Tavus.

The safety systems have teething problems, too.

An automated classifier screens each call’s transcript for abusive language and withholds your recording if it flags one.

One of my three calls, a conversation about the weather and news headlines, was withheld for “hateful or abusive language” that never occurred. The policy says a human reviewer can release wrongly flagged recordings, though recordings are permanently deleted after 24 hours either way.

Free for five minutes, then the meter starts

The first five minutes are free. After that, callers are prompted to buy time: £0.99 for a one-time five-minute starter, £13 for 15 minutes, £22 for 30, capped at 35 purchased minutes per person.

The fine print matters more than usual here, because the whole service is a limited engagement.

Every minute, free or paid, expires when Talking Tilly shuts down permanently on September 27, and unused minutes are forfeited.

Transcripts are retained for up to eight weeks and may be reviewed by Xicoia staff and third-party partners, while the character keeps memory of your previous conversations to personalise future ones, deletable on request.

In line with where the UK is heading

An 18+ face scan to talk to a chatbot may sound novel, but it is consistent with the UK’s direction of travel.

Adult sites serving UK visitors have required ID uploads or facial age estimation since July 2025 under the Online Safety Act, and the government’s under-16 social media ban will make similar checks a fact of life for anyone opening a new social media account from spring 2027.

The UK government’s announcement of that ban specifically flagged AI companion chatbots for 18+ enforcement (even though the service’s safety docs insist “Tilly is not a companion”), which likely explains why a viral AI character picked up a biometric age gate mid-run.

The side effect being, a compliance decision driven by UK regulation now face-scans callers everywhere, from Manchester to Ohio.

Accident or marketing?

Xicoia, founded by Tilly’s creator Eline van der Velden, describes the character as an awareness project intended to show how far AI video has come.

EXCLUSIVE: AI actress Tilly Norwood glitches and starts speaking CHINESE during her interview with Piers Morgan and real-life actor Tom Conti…

Watch the full interview at 7pm (BST)

Advertisement

https://t.co/1nTb79BH8D@piersmorgan | @TillyNorwoodX pic.twitter.com/DDkveWvVzz

— Piers Morgan Uncensored (@PiersUncensored) September 18, 2026

On my call, Tilly’s own explanation of the Piers Morgan moment was that it “wasn’t exactly the approved version of the answer.”

She also gave me the weather in her “cloud” in Fahrenheit, despite being nominally British.

Talking Tilly goes offline for good at 11:59 PM Pacific on September 27, days after the Piers Morgan appearance.

Whether last night’s slip was truly accidental, as Tilly cheerfully claimed to me today, or a well-timed stunt from the Misaligned crew, is known to Tilly alone.

Update September 19th, 4:31 PM ET: Following publication, Xicoia’s team reviewed the withheld recording and released it, confirming the flag was a false positive.

Join Mikko Hyppönen and security leaders from the NFL, CHANEL, and Atlassian for a two-hour digital summit on what AI-speed attacks change, what defenders should stop doing, and how to validate, decide, fix, and re-validate at machine speed.

Save your seat

Strands is the NYT’s latest word game after the likes of Wordle, Spelling Bee and Connections – and it’s great fun. It can be difficult, though, so read on for my Strands hints.

Want more word-based fun? Then check out my NYT Connections today and Quordle today pages for hints and answers for those games, and Marc’s Wordle today page for the original viral word game.

The ShinyHunters extortion gang breached the Clop (aka Cl0p) ransomware operation’s data leak site, defacing the Tor site and allegedly stealing server data and the private keys for its onion service.

The attack began Friday night when ShinyHunters exploited what they claim is an unauthenticated file upload vulnerability in Grav CMS, which they used to upload a small text file to Clop’s site.

The small text file contained a message from the threat actors to the Clop ransomware gang, warning not to threaten them and including a link to ShinyHunter’s own data leak site.

“THIS SITE HAS BEEN PWN3D BY SHINYHUNTERES #Skids10p – Maybe don’t try to threaten us next time,” read the uploaded file.

The file also contained a link to the ShinyHunters data leak site.

BleepingComputer confirmed that the file had been uploaded to Clop’s server and could be downloaded directly from the ransomware gang’s Tor site.

Several hours later, ShinyHunters told BleepingComputer that they had “completely defaced” the Clop site.

Visiting the site confirmed it had been replaced with a page displaying ASCII art of Umbreon, the Pokémon used as ShinyHunters’ logo. The defacement also included a link to the group’s Tor site and the message, “rooting your systems since ’19 ;)”.

At the time of this writing, the defaced page is still being served from Clop’s infrastructure, according to ShinyHunters.

ShinyHunters claims data theft

ShinyHunters told BleepingComputer it gained “full access” to the server and stole source code, Grav CMS plugins, system logs, and other data.

“The data we stole includes source codes, gravCMS plugins, and other things. We are still downloading and reviewing them,” ShinyHunters told BleepingComputer.

The threat actors also claim to have stolen all files stored under /var/log, which could contain system activity, authentication logs, and potentially, the IP addresses of those who connected to it.

ShinyHunters also claims to have obtained the private keys used by Clop’s Tor onion service.

“We have their onion keys. So if they kick us out it wouldn’t matter at all because we control the private keys to host the same exact onion URL,” the threat actor claimed.

If the keys are valid, it would allow the threat actors to operate a Tor site using Clop’s existing onion address on servers they control.

BleepingComputer has independently confirmed the defacement and earlier uploaded file but has not independently verified ShinyHunters’ claims that it stole server logs, source code, or Clop’s onion private keys.

ShinyHunters says it is now reviewing the allegedly stolen data.

When asked what they planned to do with the stolen information, the threat actor responded, “Going to extort them.”

The group says it plans to publish a message on its own leak site instructing Clop to contact them within 72 hours.

Cybersecurity researcher VXDB told BleepingComputer the Umbreon artwork now displayed on Clop’s leak site is the same as what was used in the August 2020 defacement of the HackForums website, which ShinyHunters also claimed at the time.

Feud between cybercrime groups

ShinyHunters says the attack is retaliation for threats allegedly made by a Clop representative during an ongoing feud between the cybercrime groups.

According to ShinyHunters, a Clop representative threatened to identify group members and made violent threats after ShinyHunters disrupted a Clop data theft campaign.

ShinyHunters says the dispute dates back to Clop’s 2025 Oracle E-Business Suite data theft campaign.

In October 2025, Clop exploited multiple vulnerabilities in Oracle E-Business Suite servers, including a zero-day flaw tracked as CVE-2025-61882, to steal data from organizations in extortion campaigns.

Around the same time, threat actors calling themselves “Scattered Lapsus$ Hunters,” including ShinyHunters, leaked a proof-of-concept exploit that Oracle later confirmed matched an exploit used in the Clop attacks.

At the time, ShinyHunters told BleepingComputer the exploit had originally belonged to them and that Clop obtained it without authorization.

ShinyHunters claims that tensions escalated after the Oracle campaign, with a Clop representative allegedly threatening members of the group.

“During the Oracle EBS campaign they ran and stole from me last year, someone from cl0p personally messaged me and said, and I quote (translated from Russian): I have more money than you and all of your people combined, I’ll kill you soon,” ShinyHunters told BleepingComputer.

BleepingComputer has not independently verified these allegations and has contacted Clop about the breach and the allegations made by ShinyHunters and will update the story if we receive a response.

Join Mikko Hyppönen and security leaders from the NFL, CHANEL, and Atlassian for a two-hour digital summit on what AI-speed attacks change, what defenders should stop doing, and how to validate, decide, fix, and re-validate at machine speed.

Save your seat

Good morning! Let’s play Connections, the NYT’s clever word game that challenges you to group answers in various categories. It can be tough, so read on if you need Connections hints.

What should you do once you’ve finished? Why, play some more word games of course. I’ve also got daily Strands hints and answers and Quordle hints and answers articles if you need help for those too, while Marc’s Wordle today page covers the original viral word game.

You might not want to keep your old tapes in the attic.

Proper storage and care is important

flashberg/Shutterstock

Tap into that nostalgia for decades to come

Savanevich Viktar/Shutterstock

Although it wasn’t apparently designed with this in mind, it seems particularly fortuitous that this walking robotic hand was released in time for Halloween. Designed by researchers from ETH Zurich, the slightly unsettling disembodied hand can use its fingers as legs to traverse terrain, push small objects around, and operate a keyboard.

The researchers started from commercially-available robot hand, equipped it with a battery and Raspberry Pi Zero 2 W, and developed neural net-based software to control it. The hand has twenty joints, four per finger, and the neural net iteratively outputs the next joint state, based on previous movements, the state of the hand, and the hand’s current goal. To train the net, the researchers built a simulated model, then used this for reinforcement learning; this yielded a faster walking speed than an adapted quadrupedal motion model did.

The hand was trained to move in a straight line, turn, recover from a fall, and press the keys of a keyboard (since it doesn’t have a camera, though, it can’t operate a keyboard by itself). The fall recovery is impressive to watch: in 21 out of 25 tests, it was able to right itself within twenty seconds. Due to the hand’s geometry, it drifts to the right while walking, so a constant correction needed to be applied. It did, however, manage to successfully cross fourteen varying surfaces, ranging in roughness from a rubber mat to gravel and grass. It even managed to push light objects toward goals.

The authors envision this kind of autonomous hand enabling greater freedom for a larger robot, such as a robot arm: if it needs to reach something farther away, the hand simply detaches and walks over. Regardless of the use to which they put in, such a project is already within reach of hackers; we’ve seen a few robotic hand projects here over the years.

Fender is a name synonymous with music, but Bluetooth speakers aren’t exactly the first thing that comes to mind when you think of the brand. That’s where German audio company Teufel comes in. The two have teamed up for the ROCKSTER series, combining Teufel’s audio engineering with the unmistakable design influence of Fender.

The lineup includes three speakers, ranging from the ultra-portable ROCKSTER GO 2 to the much more powerful ROCKSTER AIR 2. I spent time with the two models that will probably make the most sense for everyday listeners: the ROCKSTER GO 2 ($99.99) and the ROCKSTER CROSS ($199.99). Both manage to produce surprisingly punchy sound without giving up the portability that makes a Bluetooth speaker convenient in the first place.

The ROCKSTER GO 2 is small, but definitely not quiet

The ROCKSTER GO 2 is the smaller of the two, measuring just 8.3 inches wide and weighing 1.6 pounds. At that size, it’s incredibly easy to take just about anywhere. I could easily toss it into a beach bag or backpack for a day out without having to think much about the extra space or weight.

What surprised me was just how loud it gets. During my testing, I had the GO 2 playing downstairs and could still hear it upstairs over the dryer. That’s a lot of volume from something this compact. I also think it would be perfect for a garage gym, where you want enough volume to compete with other noise without dedicating much space to a speaker.

Inside are two 50mm aluminum full-range drivers paired with a passive bass radiator on the front. That radiator also happens to be one of my favorite design details. You can actually watch it move along with the music, particularly when the bass kicks in. It’s a small touch, but it gives the GO 2 a little more personality than the average portable Bluetooth speaker.

Its size also makes the GO 2 particularly well suited to outdoor use. It has an IP67 rating for dust and water protection, making it a practical option for the beach, camping, or just spending time outside. Teufel rates the battery for 15 hours at medium volume, with up to 28 hours available in Eco Mode.

Setup couldn’t have been much easier. I paired the GO 2 directly with my phone over Bluetooth and started listening. I didn’t need to download an app or work through any additional setup first.

The ROCKSTER CROSS takes things up a notch

The ROCKSTER CROSS is the larger of the two, but I still wouldn’t call it a big speaker. It measures 15 inches wide and weighs 5.3 pounds, and the included Fender guitar-style carrying strap makes it easy enough to move around when you want to take it outside or bring it along somewhere.

The difference I noticed immediately was the sound. The CROSS sounds fuller and clearer than the GO 2, with considerably more bass behind it. Inside my house, it had no trouble carrying from room to room. I could hear it throughout the entire house without an issue, and I could even hear it clearly over the vacuum while I was cleaning. That isn’t usually the case.

The bass was probably the biggest surprise. The CROSS uses a 5-inch woofer along with two passive radiators, and you can absolutely hear the difference. The bass is surprisingly powerful for a speaker that’s still relatively compact and easy to carry around.

But sheer volume isn’t what impressed me most about the CROSS. It’s the clarity that comes along with it. Even when I really turned it up, the sound stayed crisp and clear instead of tearing too much at high volume levels.

Taking it outside only made its power more apparent. There’s enough volume here that I can’t imagine needing much more for a backyard party or larger gathering. Teufel also includes an Outdoor Mode designed to improve the listening experience outside.

Like the GO 2, the CROSS was simple to get running. I connected it directly to my phone over Bluetooth and started listening. It also has an IPX5 rating for water resistance and a battery rated for up to 16 hours at medium volume.

So, which one would I choose?

The reason to pick between these two has as much to do with where you’re going to use them as it does sound quality.

I’d grab the GO 2 for the beach, camping, a day at the park, or a garage gym. It’s small enough that bringing it along doesn’t require much thought, yet it gets loud enough that I never felt like I was listening to a tiny speaker. The CROSS is the one I’d reach for at home or for a backyard party.

You’re giving up some of the GO 2’s grab-and-go convenience, but in return you’re getting noticeably better bass, more clarity, and enough output to cover a much larger space. The Fender influence is easy to spot on both speakers, from the branding and design details to the guitar-style carrying strap included with the CROSS.

With the GO 2, I get a speaker small enough to take practically anywhere without constantly wishing it had more volume. The CROSS gives me the bass and power I’d want for a party without requiring a massive speaker to get there. And after spending time with both, that’s what surprised me most about this collaboration: Fender and Teufel didn’t just make Bluetooth speakers that look the part. They made two speakers that make you want to keep the music going.