Feed aggregator

Meze Audio 105 AER: Two-minute review

Meze Audio doesn’t just stick to the super-high end with its headphone designs – it just seems that way sometimes, because its super-high end headphones are so very good. Mind you, its idea of what constitutes ‘entry level’ won’t chime with everyone – these 105 AER open-backed, wired, over-ears are $399 / £369 / AU$799 a pair.

That money does buy a nicely designed, beautifully finished pair of headphones, though certainly they’re a notch or two up from the equivalent Grados, say, where perceived value is concerned. And while the lack of a balanced cable option is a bit stingy, there’s no arguing with the lengths Meze Audio has gone to where the dynamic drivers that do the audio business are concerned.

And in pretty much every respect, there’s no arguing with the way they sound either. In every meaningful respect, the 105 AER get the job done in fine style: they’re an easy listen in the most positive way. Detail levels are high, soundstaging is solid, dynamic headroom is considerable, low-end presence is impressive... you name it, the Meze Audio understand it. Some of the best wired headphones money can buy then? Oh certainly, if you'll accept their open-backed nature.

If you’re in the market for wired, open-backed headphones – with all of the caveats those descriptions imply – and you have this sort of money to spend, it would be grossly negligent not to check the 105 AER out.

(Image credit: Future) Meze Audio 105 AER review: Price & release date

• Released December 1, 2024
• Priced: $399 / £369 / AU$799

Meze Audio isn't often seen in the sub $400 space – see the $1,999 / £1,799 / AU$3,399 Meze Audio Liric for reference, (now into their second iteration), however, the company has dipped its toes into more consumer-accessible pools recently, perhaps most notably with its $159 Alba in-ears.

At this price, Meze knows it could turn some heads – most pertinently from music lovers used to seeing such fees from Bose (whose QuietComfort Ultra cans are just a bit pricier) and Sony's top-tier wireless designs. Will the bid win them over? It depends – wireless audio has, and will always have, its perks. But on audio quality alone, there really is no contest…

(Image credit: Future) Meze Audio 105 AER review: Specs

(Image credit: Future) Meze Audio 105 AER review: Features

• 50mm full-range dynamic drivers
• 5Hz - 30kHz frequency response
• 112bB sensitivity, 42 ohms impedance

Just like every other pair of wired over-ear headphones, the Meze Audio 105 AER are not exactly overburdened with features. But just like every other pair of Meze Audio headphones I’ve tested, the 105 AER features are very thoroughly implemented.

Let’s be honest, though; when I talk about ‘features’, really I’m talking about the drivers that deliver sound to your ears. Everything else about the 105 AER is more correctly found in the ‘design’ section – so let’s talk about the drivers, shall we?

The ‘feature’ here is a 50mm full-range dynamic driver – or, more correctly, two of them. It’s closely based on the driver Meze Audio fits to its considerably more expensive models, with some minor adjustments to the frame and driver membrane to keep weight (and costs) down.

The ‘W’-shaped dome is made of a carbon-fibre/cellulose composite. Meze Audio prizes it for its durability and light weight, and reckons it rejects resonances more effectively than any alternative material. The torus – the ring that surrounds the dome – is made of semicrystalline polymer. It too is light and responsive, and its impressive damping characteristics keep vibrations and resonances to a minimum. Finally, there’s a copper/zinc alloy stabiliser around the outside of the membrane to further reduce distortion.

This arrangement results in a frequency response of 5Hz - 30kHz, manageable impedance of 42 ohms, and helpful 112dB sensitivity. So while the 105 AER are light on features, they are demonstrably fit for purpose.

• Features score: 5/5

(Image credit: Future) Meze Audio 105 AER review: Sound quality

• Impressive powers of resolution
• Big, unified and poised sound
• Great insight and variation

One of the major benefits of the open-backed arrangement, in theory at least, is a spacious and open presentation of music. The Meze Audio 105 AER are one open-backed model that validate this theory.

A 2.8MHz DSD file of Radiohead’s Reckoner delivered by the 3.5mm output of a FiiO M15S digital audio player proves the point in some style. This is a complex, element-heavy recording that modulates through a number of dynamic variations, but throughout it the 105 AER maintain a big, well-defined and properly organised soundstage on which every individual element gets plenty of space in which to express itself. The resolution of the stage is straightforwardly impressive, and even though the Meze Audio offer plenty of separation they nonetheless present the record as a singular, unified whole. There’s a coherence to the way the song is delivered that makes it seem of a whole, like a performance.

Detail levels are high across the board. The 105 AER load on the information at every point, but are especially adept at offering lots of variation at the bottom of the frequency range. The top end is substantial and detailed, sure, and the midrange is articulate too… but where bass is concerned, the Meze Audio are able to offer a great deal of textural and tonal variety while still maintaining straight-edged control and ample weight.

The whole frequency range hangs together nicely, and there’s a smooth transition from top to bottom – the 105 AER play no favourites and transitions from lowest to highest frequencies evenly. The tonal balance is quite carefully neutral, which allows recordings to reveal their balance without the headphones sticking their oar in too obviously. So a 16bit/44.1kHz file of Aretha Franklin’s How I Got Over is as warm as an 18 tog duvet, while similarly sized file of Kraftwerk’s Europe Endless has just the right sort of austerity.

The Meze Audio handle the dynamics of harmonic variation with just as much confidence and positivity as they do the dynamic shifts in volume or intensity. They offer convincing rhythmic expression and a naturalistic way with tempo management. In fact, they have the sort of direct, unequivocal overall personality that makes every listen an event and makes every recording sound like it deserves your attention.

• Sound quality score: 5/5

(Image credit: Future) Meze Audio 105 AER review: Design

• 1.8m detachable cable with 3.5mm termination
• Self-adjusting headband
• Detachable velour earpads

I’ve previously described the design of some of Meze Audio’s pricier over-ear headphones as ‘overwrought’ – and at the time I thought I was being quite kind. So it’s nice to be able to report that having to keep costs down has resulted in a design that makes the 105 AER look coherent, quite elegant, and definitely no more than ‘wrought’.

The detachable ear pads are of velour-covered memory foam. They feel nice, sit comfortably without warming your head too quickly, and can easily be cleaned. The almost-semicircular outer headband that connects the two earcups is of slender stamped magnesium and the inner headband that’s the actual contact point is self-adjusting and made of PU leather. The outer part of the earcups is an interestingly organic design and made from ABS-PC thermoplastic. The result is Meze Audio’s lightest over-ear headphone to date – a very manageable 336g.

Each earcup needs wiring, and the 105 AER are supplied with a 1.8m length of braided cable that has the necessary pair of 3.5mm connections at one end and a single 3.5mm connection at the other. There’s also a 6.3mm adapter included. It’s a pity there’s no option of a balanced cable, but perhaps if we all pester Meze Audio then something might be forthcoming.

The 105 AER are supplied with a hard, zip-fastening travel case that’s a cut or two above the alternatives supplied by rival brands with their $399-ish over-ear headphones. A little pouch inside keeps your cable tangle-free when the headphones are not in use – and there’s enough room in there for a second cable, no problem. Just saying…

• Design score: 5/5

(Image credit: Future) Meze Audio 105 AER review: Value

• Accomplished audio performance
• Convincing standard of build and finish
• Comfortable fit, coherent looks

It’s not even a question of ‘value’, really – but about the only way the Meze Audio 105 AER don’t make a strong-going-on-compelling case for themselves concerns the way they fit. Or, more correctly, the number of people they will fit – because although that headband is self-adjusting, it has upper and lower limits, and the lower limit isn’t all that low. In every other respect, though, it’s hard to make a justifiable complaint about the value that’s on offer here.

• Value score: 4.5/5

Should I buy the Meze Audio 105 AER?

(Image credit: Future) Buy them if...

You value balanced, insightful and organised sound
That description could conceivably be interpreted to mean the 105 AER are not much fun - but that’s not the case. It’s just that ‘fun’ isn’t all they are…

You do your listening alone
Open-backed headphones are quite anti-social at the best of times, and the 105 AER leak sound in the manner of a sonic colander

Your source(s) have unbalanced headphone sockets
A 3.5mm connection with a 6.3mm adapter is all well and good, but what about those of us whose source equipment has a balanced output too?

Don't buy them if...

You’re after outright low-frequency wallop
The 105 AER are a balanced and nicely poised listen - which means low frequencies are given the correct sort of emphasis rather than being overstated

Your head is on the small side
The headband is self-adjusting, sure - but only up to a point. The more petite-headed among us may find it tricky to get a satisfactory fit

Meze Audio 105 AER review: Also consider

Sennheiser HD-660S2
The Sennheiser HD 660S2 look pretty purposeful where the Meze Audio 105 AER look relatively dainty, but the sound they make is balanced, poised very carefully neutral – and they come with a 4.4mm cable.
Read our Sennheiser HD-660S2 review for more

Grado Hemp
Grado's Reference Series Hemp may cost a little more but they’re very accomplished performers, with a botanical point of difference – it doesn’t prevent them looking like something the rear gunner might have worn during a bombing raid in 1944, mind you…
See our Grado Hemp review for the full story

How I tested the Meze Audio 105 AER

• Connected to an Apple MacBook Pro, a Naim Uniti Star and a FiiO M15S
• With a whole lot of different styles of music in a number of different formats
• Alone, because I’m not a barbarian

Because they’re open-backed, the 105 AER aren’t really candidates for mobile use. So all my listening took place either at my desk, connected to my laptop or to a digital audio player (the 4.4mm balanced output of which went sadly unused), or in my listening space connected to a Naim Uniti Streamer. At the desk, listening consisted of digital audio files of various types and sizes, while the connection to the Naim allowed for vinyl and CD content as well as network-attached stuff. And this happened, on and off, for well over a week…

• First reviewed: February 2025
• Read more about how we test

With a more crypto-friendly White House, here is what experts say you can expect from bitcoin and other digital currencies this year.

Eating organic is normally associated with tote-wielding liberals, but RFK Jr. is shifting the politics of food in unusual ways.

An NPR listener wonders whether her husband's relationship with a female colleague is too close for comfort. He says she's being competitive for no reason, and that he sees the woman as a 'sister.'

(Image credit: Eoneren/Getty Images)

The Sims, at Electronic Arts in Emeryville, California, on March 24, 2008.'/>

The original launched in February 2000 and spawned one of the most successful franchises in video game history. NPR's Susan Stone went hands-on with The Sims shortly after its debut.

(Image credit: Ryan Anson)

Trump has shown an affinity with many of the little guys — what he called in 2017 "the forgotten men and women." But he also has shown an affinity with some of the fattest cats of all.

(Image credit: Shawn Thew)

Siegel is the first American to be released as part of this deal. A total of 18 hostages have been released since the ceasefire took effect. Hundreds Palestinians have been released so far.

(Image credit: AFP)

If your partner, friend or situationship likes working out, takes their recovery seriously or loves the outdoors, these are some gifts they'll love.

I’ve never been a huge fan of ultra-wide cameras, even on some of the best camera phones. Too often, the cameras don’t have a high enough resolution to produce crisp photos when scrutinizing the detail, or there’s too much warping on the edges. I’ve always felt the physical limitation of the small sensors and lenses used in smartphone ultra-wide cameras hold them back from delivering the more impressive shots of main and telephoto cameras.

I’ve also never been a fan of how Samsung’s ultra-wide cameras in its Galaxy S-series, with photos that appear noisy and lacking sharpness, even occasionally looking a bit smudged when peering deeper into a shot.

However, with the Samsung Galaxy S25 Ultra, the South Korean tech giant appears to have turned a corner with its ultra-wide cameras. That’s because the S25 Ultra comes with a 50-megapixel rather than the 12MP ultra-wide camera of the Galaxy S24 Ultra.

It’s been a common idea that megapixels don’t mean a great deal when it comes to smartphone cameras, especially those on the best phones. But I’m starting to see a shift in the sands in that there’s only so much smart computational photography can do with 12MP sensors and images. In contrast, more detail sucked in by a higher megapixel sensor effectively means more data to process and, thus, more scope to get a better photo.

With the caveat that others on the TechRadar team and I are continuing to test the Galaxy S25 suite of smartphones, I quickly pitched the ultra-wide camera of the Galaxy S25 Ultra against the 12MP ultra-wide on my iPhone 16 Pro Max; my main phone.

In the photo samples below, you’ll see the Galaxy S25 Ultra’s ultra-wide camera in its default 12MP mode – an annoying default, but I understand it saves on file size – and 50MP shots, alongside 12MP iPhone 16 Pro Max snaps.

Samsung Galaxy S25 Ultra vs iPhone 16 Pro Max ultra-wide photos Image 1 of 4

Galaxy S25 Ultra 50MP ultra-wide (Image credit: Future)Image 2 of 4

iPhone 16 Pro Max 12MP ultra-wide (Image credit: Future)Image 3 of 4

Galaxy S25 Ultra 50MP ultra-wide (Image credit: Future)Image 4 of 4

(Image credit: Future)

At a quick glance, it’s not particularly easy to see huge differences between the shots other than different interpretations of colors; Samsung, as usual, favors punchier colors over the iPhone’s more muted take.

But look closer, and you’ll see finer details in the foreground of the Galaxy S25 Ultra’s 50MP ultra-wide shots vs the iPhone 16 Pro Max’s. Punch into a shot, and the Galaxy S25 Ultra resolves sharper details than its rival.

The same is broadly true when comparing the 12MP ultra-wide versus the 50MP shots on the Galaxy S25 Ultra. You have to go on the hunt for changes, as both shots use the same camera sensors; just the 12MP images are being produced with pixel-binning to make a 50MP snap smaller.

Still, from what I can tell, this process naturally loses some detail on a shot but also removes a degree of dynamic range; in the 12MP shots, shadows and highlights are more uniform, whereas in the 50MP photos, there’s more of a gradient from light to dark.

Samsung Galaxy S25 Ultra 50MP vs 12MP ultra-wide photos Image 1 of 4

Galaxy S25 Ultra 50MP ultra-wide (Image credit: Future)Image 2 of 4

Galaxy S25 Ultra 12MP ultra-wide (Image credit: Future)Image 3 of 4

Galaxy S25 Ultra 50MP ultra-wide (Image credit: Future)Image 4 of 4

Galaxy S25 Ultra 12MP ultra-wide (Image credit: Future)

I find this pertinent, as recent conversations about flagship phone cameras have focused on how they tend to produce somewhat flat photos by bringing up the shadows and pulling down bright parts. This can bring out more detail but results in a photo that doesn't look quite true to life and ends up lacking some contrast that can help make a photo pop.

So, my key takeaway for anyone planning to get a Galaxy S25 Ultra is to use the 50MP mode in the ultra-wide camera for the best results.

And I think this is a sign that Apple should look to follow the example set by the Galaxy S25 Ultra and the Google Pixel 9 Pro, as well as the Pixel 8 Pro, and adopt a larger pixel count with the ultra-wide camera on the iPhone 17 Pro.

Head to the comment and let me know which ultra-wide photos you prefer, and do check out our full Samsung Galaxy S25 Ultra review; also, keep TechRadar bookmarked as we’ll have more thoughts and experiences with the Galaxy S25 phones as we keep digging into them.

You might also like

• iPhone vs Galaxy video: which would you choose?
• Check out our hands-on Samsung Galaxy S25 review
• And read our hands-on Samsung Galaxy S25 Plus review

This week was wild. DeepSeek came and changed the AI landscape, but it seems like the full impact of its arrival is still yet to be fully understood. Meanwhile Garmin accidentally bricked many of its devices globally thanks to an errant update, and Android XR broke cover in our first hands-on look at the software.

Al that and more of the otehr big tech news stories you might have missed from the week are recapped below so you can catch up with everything then get on with your day.

If you're looking to catch up on the biggest entertainment of the moment too then chekc out our picks for the 7 new movies and TV shows to stream this weekend (January 31).

7. DeepSeek R1 conquered the AI world

(Image credit: Future)

It’s been a rollercoaster week for the AI industry with the DeepSeek app rocketing to the top of the Apple App Store and beating ChatGPT. DeepSeek is a Chinese startup who claim to have developed their deep reasoning model, R1, for a fraction of the price that US AI companies have invested in their technologies.

It also offers people access to it for free in its chatbot and at a much reduced price as an API compared to OpenAI. The news that a Chinese company can produce an equivalent technology for a fraction of the price caused a complete crash of the share value of US tech companies that are heavily invested in AI. Nvidia came off particularly badly with $600 billion being wiped off their stock value in just one day. Share prices have since recovered, but it doesn’t feel like the AI market will ever be the same again.

• Read more: DeepSeek live – all the latest news

6. But did it cheat to win?

(Image credit: Getty Images)

Questions are already being asked about how DeepSeek managed to produce its AI model so quickly, especially after OpenAI claimed it had evidence of distillation of some of its models. Distillation is a development technique where you piggyback off another model’s learning, but is against the OpenAI terms and conditions of use. To make matters worse for DeepSeek, it appears that when questioned about which AI model is best, DeepSeek will often refer to itself as ChatGPT, which some users think of as a bit of a smoking gun.

OpenAI released a statement on the matter, stating: “We know that groups in the PRC are actively working to use methods, including what’s known as distillation, to try to replicate advanced U.S. AI models. We are aware of and reviewing indications that DeepSeek may have inappropriately distilled our models, and will share information as we know more. We take aggressive, proactive countermeasures to protect our technology and will continue working closely with the U.S. government to protect the most capable models being built here.”

• Read more: DeepSeek just insisted it's ChatGPT, and I think that's all the proof I need

5. Garmin became the Sonos of wearables

Garmin's blue triangle of death (Image credit: Angela MacAusland)

A tough few months of software quirks boiled over into a massive Garmin outage this week that saw users in their thousands completely lose access to their devices. A rogue GPS file accidentally pushed to Garmins saw user devices plunged into a boot loop dubbed the 'Blue Triangle of Death.' Few customers found they could escape, and even those who did encountered further issues with connectivity and synchronisation.

The outage lasted well over 24 hours and we heard from hundreds of customers, owners of devices like the Forerunner and Epix range, as well as more niche devices like Garmin's dive computers, cycling computers, and Approach Golf Range.

Garmin has since fixed the underlying issue, but lots of customers remain stuck in boot loops and without a device. Plenty have even told us they regret their Garmin purchases or plan to depart for companies like Apple. Garmin has a PR disaster on its hands that has vastly overshadowed the launch of the Instinct 3. The company has lots of work to do, or it risks becoming the Sonos of wearables.

• Read more: Garmin outage FIXED: What to do if your device still isn't working

4. Nvidia’s new RTX 5090 came and went

Nvidia RTX 5090 (Image credit: Future)

Nvidia’s new graphics cards are here and… they’re already gone as stock sold out lightning fast across the internet for these highly sought after GPUs. Some will have been bought by legitimate gamers looking to upgrade their rig – and they’re in for a treat with our RTX 5090 review highlighting how it’s a seriously powerful upgrade that approaches complete overkill performance, and our RTX 5080 review telling us it boasts some solid gains too at a more modest (but still not inexpensive) price.

Unfortunately some Nvidia cards also seem to have been picked up by scalpers taking advantage of the limited availability to flip them for a profit on resale sites by listing the GPUs for several times their MSRP. Resale sites have also been flooded with listings that seem to be selling the new GPUs for prices closer to what they should cost, but when you take a closer look you’ll see they’re promising to send you a picture of the graphics card you desire rather than the real thing.

So our advice is to wait for another official drop so you don’t get tricked by a scammer or pay three or four times more for the GPU you’re after – and you can follow our coverage for updates on when and where these drops will happen.

• Read more: Where to buy Nvidia RTX 5090 GPUs in the US and UK

3. Nothing Phone 3 was teased

Nothing Phone 2 (Image credit: Future | Alex Walker-Todd)

Nothing shared a short video of its CEO Carl Pei this week. Normally that wouldn’t be super newsworthy, but as the camera zooms in on his glasses we see the words “Launch 4 March” reflected in the lens. A previous Nothing Phone 3 teaser suggested this would be the announcement date for its next smartphone, but now it’s all but certain.

What’s more interesting is a cut in the middle of the clip that seems to shift from a shot from a wide lens, to one from a telephoto lens – the latter of which is absent from all Nothing phones. We’re taking this as a hint that the Nothing Phone 3 could have a triple-lens camera – with a telephoto camera joining the wide and ultra-wide lens setup found on the Nothing Phone 2.

• Read more: Nothing hints that the Phone 3 could have the camera feature it needed

2. We got our first in-depth look at Android XR

Project Moohan (Image credit: Lance Ulanoff / Future)

Thanks to a hands-on video from Marques Brownlee on his YouTube channel, MKBHD, we got our first look at Android XR and the Project Moohan prototype this week. While many details including price, release, and some specs are still a mystery we did get to see aspects of the headset’s design and software.

It looks a lot like a combination of the Meta Quest Pro and Apple Vision Pro. There’s no over-the-head strap but instead a Meta Quest Pro-like adjustable strap round the back, and the Android XR headset borrows the Quest’s optional light blockers to change between a VR-optimized and MR-optimized design. It also uses an external battery pack like the Vision Pro to help save weight.

As for software, Brownlee’s video predominantly focused on Gemini’s integration – which looks impressive. Though many of the AI features look like they’d be much better suited to a pair of AR glasses – so while we expect Moohan will be impressive, we’re already waiting to see what XR hardware Samsung and Google will launch next.

• Read more: Android XR comes out swinging with our first look at Samsung Project Moohan

1. Our Friendly Neighborhood Spider-Man swung in

Spider-Man from Your Friendly Neighborhood Spider-Man (Image credit: Marvel Animation/Disney Plus)

Our favorite wall-crawling superhero is finally back! The latest adaptation of Peter Parker to enter the fray is in Marvel’s new animated series Your Friendly Neighborhood Spider-Man.

The show has a unique release schedule, though, as each week two episodes will be released instead of one – the first two of which debuted on Disney Plus this Wednesday (January 29).

TechRadar’s senior entertainment reporter Tom Power has already seen all 10 episodes and says the show’s “a spectacularly fun ride” in his Your Friendly Neighborhood Spider-Man review.

Which is great news for Spidey fans, because two more seasons have been greenlit and the show’s creator already has “big ideas” for where it could go next.

• Read more: Every Marvel Easter egg and reference in Your Friendly Neighborhood Spider-Man episodes 1 and 2: Nico Minoru, Lonnie Lincoln, and more

Here are the answers for The New York Times Mini Crossword for Feb. 1.

• Nvidia's fastest GPU is unsurprisingly expensive and power hungry, just like its predecessor
• It is also extremely fast and leaves the RTX 4090 and AMD's fastest GPU, well behind
• Early benchmarks show there's far more potential once drivers and updates are rolled out

Nvidia’s latest flagship GPU, the GeForce RTX 5090, was announced at CES 2025 and has just gone on sale, although Nvidia has warned that it expects to sell out quickly. Built on the Blackwell architecture, the RTX 5090 is a successor to the RTX 4090, and features 32GB of VRAM, increased CUDA core counts, and improved memory bandwidth.

Puget Systems, which previously tested the GeForce RTX 4090, has just benchmarked the RTX 5090 across various creative applications, including Adobe Premiere Pro, After Effects, and DaVinci Resolve and as you might expect, the results were impressive.

In Premiere Pro, Puget Systems found the RTX 5090 was slightly faster than the RTX 4090 by about 9% in terms of overall performance. In Blackmagic’s DaVinci Resolve Studio, the GPU had a 17% lead over the 4090 and 35% over the 3090 Ti. Puget noted however, that the “5090 was run on a slightly different version of Resolve than the rest of the cards - a review version designed to be fully compatible with the 50-series card and which we expect to be incorporated into the application in the near future.”

Obliterating AMD

(Image credit: Puget Systems)

Adobe After Effects also benefits from the increased power of the RTX 5090.

According to Puget Systems, the GPU recorded a "massive 35% gen-on-gen improvement over the RTX 4090" in 3D rendering tasks, making it a solid choice for motion graphics professionals. Elsewhere, Unreal Engine benchmarks suggest the RTX 5090 leads the RTX 4090 by 17% overall. In rendering applications such as Blender and V-Ray, "the RTX 5090 is a massive 38% faster than the 4090 and three times as fast as the 3090 Ti."

The RTX 5090 outperformed AMD’s fastest consumer GPU, the Radeon RX 7900 XTX, across multiple tests, although AMD’s card faced some challenges. In Adobe After Effects, Puget Systems observed that the 7900 XTX "currently struggles with the 'Advanced 3D' renderer, with the 7900 XTX half as fast as even the 2080 Ti." It was a similar story in Unreal Engine where "AMD struggles with ray tracing, seeing a much larger performance drop than Nvidia when enabling the feature."

Despite the solid benchmarks, there are some early software compatibility issues. Puget Systems noted that "at present, the RTX 5090 is not supported in either Redshift (Cinebench) nor Octanebench" and has known performance problems in V-Ray’s CUDA rendering. Nvidia is expected to resolve these issues with future driver updates.

With a launch price of $2,000, the RTX 5090 is obviously positioned as a premium option for users who need the highest level of performance. As Puget Systems concludes, "if you need the most powerful consumer GPU ever made, this is it."

You might also like

• These are the best video cards for video editing you can buy right now
• Nvidia is preparing for the post-GPU AI era
• Geforce GPUs beat the competition in this popular video software benchmark

What ISPs offer the best coverage, speed and pricing in Phoenix? See what broadband options you can get in the Valley.

• Similiarweb data claims DeepSeek’s usage doubled to 12 million people worldwide in the space of two days
• Release of DeepSeek’s R1 model has prompted panic in Western AI leaders worried about their bottom line
• ChatGPT continues to outperform DeepSeek in sheer page hits, but not in growth

In the space of two weeks, open source and MIT-licenced Chinese large language model (LLM) DeepSeek has taken the AI tool world by storm, sending Western AI-leader Nvidia stock plummeting and prompting OpenAI’s Sam Altman to accuse DeepSeek’s developers of using its models to train theirs.

Western AI figureheads are right to be on their toes, as new data shared exclusively with TechRadar Pro from Similarweb has shown DeepSeek’s centralised web and mobile app version (the nature of open source means that users can run various models locally on their own hardware, which Similarweb would not have data for) is seeing considerable growth.

ChatGPT versus DeepSeek

Let’s zero in on late January, as that’s when DeepSeek’s new, advanced ‘R1’ model was released. Between January 24 and January 26 2025, worldwide daily visits to DeepSeek doubled from 6.2 million to 12.4 million.

Blips in DeepSeek’s page traffic did come in the week before the model’s release, with a pronounced drop of 900,000 page views between January 15 and 18. Since January 19 (the day before the model’s release), however, the service saw steady, albeit inconsistent growth, culminating in that two-day surge; the latest data we have.

ChatGPT, meanwhile, has seen precipitous drops in page traffic before and during the release period for R1, indicating it may have already become old-hat in the eyes of many with their eye on the LLM space without DeepSeek entering the fray. The service lost 43.1 million views between January 15-18, while the biggest fall post-R1’s release came between January 23-25, with a loss of 41.3 million views.

It’s possible these are natural ebbs and flows, and that ChatGPT is bound to see bigger losses because it’s a larger operation that has been in the public consciousness for longer. It’s also important to note, although ChatGPT has seen these recent drops, the losses still amount to four times the amount of views that DeepSeek has amassed according to the latest SimilarWeb data. ChatGPT is hardly ‘dying’, either; it still managed a strong peak of 140.6 million views on January 23, three days after the release of DeepSeek R1.

The main worry, then, is growth; ChatGPT seems to have run out of it; amassing an average of 126.9 million page views in the week of DeepSeek’s latest model release, and only being able to achieve sporadic daily peaks of around 140 million views over non-consecutive days in that period. Both of these figures don’t represent growth over previous months according to the data. A glance at the available SimilarWeb figures tells me that ChatGPT was pulling in comparable numbers as late as December 2024 and as early as September.

DeepSeek may have only amassed a mean average of 7.45 million views in the same period, but that two-day doubling will concern interested competitors. The questions in play, that we just don’t know the answer to yet, are ‘how long will this rate of growth continue’ and ‘can DeepSeek become a meaningful long-term competitor in AI’? While anyone keen on competition in the space (specifically from a model able to be hosted locally) might be hopeful about the latter, the fact remains ChatGPT is a juggernaut, and it’ll be no easy feat.

What is DeepSeek?

In case you missed it, DeepSeek is a Chinese company that has managed to train R1, a model “on par with ChatGPT” and Copilot for $6 million (£4.8 million), a fraction of the cost of the $78 million] spent by OpenAI in training its latest GPT-4, all while the country endures an embargo of powerful high-end graphical processing units (GPUs) from the West.

DeepSeek began as a startup in May 2023, with founder Liang Wenfeng putting his stockpiling of Nvidia GPUs down to “curiosity” as opposed to business acumen in 2021 - before former US President Joe Biden introduced export sanctions on semiconductors to China in October 2022.

For SMBs, DeepSeek R1 is freely available on Github, and, being MIT-licensed, represents a significant opportunity for low-cost AI chatbot implementation. It’s a powerful model that, unlike ChatGPT or Copilot, can be run locally, and on modest hardware.

While rumbles of data leaks have emerged surrounding the web and Android app versions, it’s important to note that running the model yourself allows for sidestepping these concerns.

You might also like

• AI agents are proving remarkably popular - but firms still face a host of challenges
• These are the best business VPNs
• Could clever hardware hack be behind DeepSeek's groundbreaking AI efficiency?

• Squarespace teams up with Barry Keoghan for new Super Bowl ad campaign
• A comedy teaser video is available to watch
• The ad will air between the 1st and 2nd quarters at Super Bowl LIX

When I woke up this morning I didn’t think I’d be talking about bath water drinking actors, donkeys, and one of the world’s most watched sporting events - but here I am - thanks for making my day a little weirder Squarespace.

Earlier this month Squarespace, one of the best website builders, released a short teaser from the company's upcoming Super Bowl 2025 ad campaign. It offered few details beyond an unidentifiable man riding a donkey down a country track.

Yesterday, everything became a little clearer after I received a press release from Squarespace detailing some exciting new details about the ‘A Tale As Old As Websites’ Super Bowl ad campaign.

It is now revealed that the mystery man is Barry Keoghan, the Irish actor, perhaps best known for that scene in Saltburn. In this ad campaign Barry ditches his questionable methods of hydration in favor of a furry four legged friend.

Although the actual ad remains a closely guarded secret, the company has released a comedic short film titled Barry and Moseley to help build anticipation. The clip includes behind-the-scenes footage of Barry along with his co-star, discussing the donkey’s career prospects and creating a website using the Squarespace website builder.

Squarespace’s Super Bowl 2025 campaign

This is the 11th Super Bowl appearance from Squarespace. With costs for a 30-second spot reportedly sky rocketing from $7 million in 2024 to $8 million in 2025, it is clear that the company continues to see the value of this insanely expensive ad spot.

Squarespace will be hoping to continue a string of Super Bowl ad successes following its 2024 Super Bowl ad, directed by and starring legendary filmmaker Martin Scorsese, which was seen by a share of a predicted audience of 210 million.

We heard from Nate Skinner, Senior Director of Brand and Creative Strategy at Squarespace:

"In today’s media landscape, the Super Bowl remains one of the rare opportunities to captivate millions at once. It’s more than a moment to advertise - it’s a chance to share a story that resonates with a massive and diverse audience. To capitalize, you have to entertain and find creative ways to stand out.” Nate shared “By teaming up with the incredibly talented Barry Keoghan and crafting a humorous tale set in a bygone era, we’ve created a campaign that’s designed to inspire entrepreneurs to dream big and take their ideas to the next level."

The new ad was directed by Steve Rogers and will be run between the 1st and 2nd quarters. It will be shown alongside ads from other big businesses including website building and hosting company GoDaddy, Budweiser, Hellmann’s, and Uber Eats.

More from TechRadar Pro

• What time is the super bowl? We have the answer
• Wix's new AI tool aims to take you from idea to profit in record time
• Running a small business? We have listed the best small business website builders

The decision came in response to a petition by attorneys general in 22 states and Washington, D.C., seeking to block the administration's efforts to freeze payments for grants and other programs.

(Image credit: Roberto Schmidt)

OpenAI showed off its first AI Agent, Operator, last week, but it already has a scrappy competitor offering an AI tool called Browser Use that can complete tasks online for you. This Computer-Using Agent (CUA) can write, search, click buttons, and copy information from websites without you needing to touch the mouse or keyboard and without the $200-a-month ChatGPT Pro subscription.

Browser Use is actually free, at least if you're willing and able to spend some time playing with API code. I'm not very code-literate, but I naively thought I knew enough of how GitHub works to use the API version. Hours of sifting through documentation, tweaking settings, and watching examples later, I decided this would need a deeper level of coding knowledge than I have, let alone the average person browsing the web.

Happily, for me, Browser Use just debuted a cloud version that employs OpenAI's own GPT-4o model. It cuts out a lot of the heavy technical lifting and streamlines things into a more familiar chat format without any extra work. It has its limitations and costs $30, but after my inept API mess, it felt like a bargain. And even in this (still obviously unfinished) form, you still need to put some effort into engineering prompts and negotiating how the AI functions. The most limiting aspect is that you can only issue one prompt before having to start a new interaction. Despite the text box, you can’t respond to what the AI does and refine your request.

Buying AI

(Image credit: Screenshots from Browser Use)

With everything set up, I put Browser Use through a few real-world tests. First up was a price comparison task. I entered the prompt: "Navigate to Amazon, Best Buy, and Walmart and search for 'MacBook Air M2'. Extract the product name, price, and stock availability from the first five results on each site. Compare the prices and identify the lowest one. If discounts or coupons are present, record them. Provide a final summary with the best deal and where to buy it."

It did the job well, though it didn’t find any hidden discounts or coupons. Still, the fact that I could automate price tracking across multiple sites was pretty exciting. That said, a continuing issue for any agent like this comes when a website wants to check that you're human. Browser Use has a button that lets you take over whenever you want, but it will also alert you when there's a need. You can prove your humanity and then hit resume to let the AI take over again.

(Image credit: Screenshots from Browser Use) Fly AI

(Image credit: Screenshots from Browser Use)

Next came a travel planning task with the prompt: "Search for a round-trip flight from New York to London on Dec 15, 2025 on British Air. Select the cheapest option and extract details, including price, airline, and departure time."

Browser Use delivered, pulling up a British Airways flight at $750, complete with departure time and other relevant details. This could be incredibly useful for people who book a lot of travel, especially if you automate it to check for price drops regularly.

Fair weather AI friend

(Image credit: Screenshots from Browser Use)

Finally, I tested out weather prediction and planning with the prompt: “Check the 7-day weather forecast for New York City on weather.com and summarize temperature trends, rain chances, and any severe weather warnings and then suggest how to dress for it.”

Weather is one of the most popular uses for voice assistants, so I wanted to see how the AI handled a more complex request in that vein. It did very well, not only extracting the information from the forecast but suggesting which days to wear a light coat and which days I should “insulate with a warm coat and scarf, as it will be chilly with low rain chance.”

Power trip

The key difference between the two is accessibility. Browser Use is like a Swiss Army knife for developers. It has the flexibility to do almost anything within a browser, but you need to know how to use the tools. You can dig into the code, tweak it, and mold it to your exact needs. If a feature is missing, nothing’s stopping you from adding it. Browser Use, being open-source, also has an active developer community constantly refining it. That means if you run into issues, there are forums and GitHub discussions where you can likely find answers.

OpenAI’s Operator, on the other hand, is like hiring a butler. It does a lot for you but within certain constraints. Operator’s strength is its integration with OpenAI’s broader AI ecosystem, giving it access to proprietary models that can make more nuanced decisions. However, you’re locked into OpenAI’s pricing structure and limited customization options.

Browser Use isn’t perfect. Even its cloud version demands some patience. You need to craft your prompts carefully, brace yourself for troubleshooting, and occasionally start over. The cloud version may make up for some of this later, but for now, the limits of not being able to edit or respond within the conversation put hard limits on its otherwise flexible nature.

And the speed can be frustrating as well. Check out a video of my second test; this is four times the speed of the actual process.

Right now, Browser Use is best suited for people who enjoy tinkering, such as developers, researchers, and automation geeks who don’t mind getting their hands dirty. If you’re willing to put in the effort, you’ll get a powerful, flexible tool that costs way less than its competition.

But if you’d rather not spend your weekend wrestling with configuration files, Operator may be the more forgiving option. Either way, web automation is ready for a boom.

You might also like

• OpenAI's first AI Agent is here, and Operator can make a dinner reservation and complete other tasks on the web for you
• I discovered a surprising difference between DeepSeek and ChatGPT search capabilities
• ChatGPT Tasks can start taking over your calendar and remind you to finish your to-do list

Here's everywhere you can hail a ride, and where else Waymo will soon arrive.

A small plane crashed in Northeast Philadelphia 30 seconds after it took off, Pennsylvania's governor said Friday. The Federal Aviation Administration said two people were aboard.

(Image credit: Matt Rourke)

Clinical trials may have to be scrapped, research applications will be pushed back, and unpaid researchers will quickly leave the sector—even if the Trump administration’s funding pause is only temporary.