Poolside, the San Francisco AI lab that has spent most of its three-year existence quietly selling coding models to governments and defense agencies, released its most capable model to date on Tuesday — and made an unusually aggressive bet that radical transparency, not raw scale, is how a smaller lab competes at the frontier.

The model, Laguna S 2.1, is a 118-billion-parameter Mixture-of-Experts (MoE) system that activates only 8 billion parameters per token, supports a context window of up to 1 million tokens, and — according to benchmarks published by the company — matches or beats open models several times its size on agentic coding tasks. The weights are available immediately on Hugging Face under the permissive OpenMDW-1.1 license.

The headline numbers are striking for a model this small. Poolside reports that Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, a benchmark of long-horizon terminal tasks, placing it 11th on the company’s compiled leaderboard — ahead of DeepSeek-V4-Pro-Max, a 1.6-trillion-parameter model that scored 64.0; Thinking Machines’ 975-billion-parameter Inkling, at 63.8; and Nvidia’s 550-billion-parameter Nemotron 3 Ultra, at 56.4. On SWE-Bench Multilingual, it posts 78.5%, and on SWE-Bench Pro‘s public dataset, 59.4%.

Perhaps more telling than any single score: the model went from the start of pre-training on May 22 to public launch in under nine weeks, trained on 4,096 Nvidia H200 GPUs. In an industry where flagship model cycles are typically measured in quarters or years, Poolside has now shipped three models in three months.

Why the West’s open-weight AI gap has become a boardroom issue

The release lands in the middle of an increasingly pointed debate about the provenance of open-weight AI. Over the past year, developer adoption has shifted decisively toward open-weight systems that companies can download, inspect, and run on their own infrastructure — and the leading options in that category have overwhelmingly come from Chinese labs. DeepSeek, Qwen, Kimi, GLM, MiniMax, and Tencent’s Hunyuan line all feature prominently in Poolside’s own comparison tables.

Poolside’s accompanying press release frames Laguna S 2.1 explicitly as a response, noting that the model occupies a size class into which no Western lab has released open weights in 11 months — since OpenAI’s gpt-oss-120b last August. “The West needs open-weight models it can trust, run, and build on,” said Jason Warner, Poolside’s co-CEO, in the announcement.

Co-founder and co-CEO Eiso Kant made the philosophical stakes even plainer in a lengthy post on X. “I believe intelligence should and will become a commodity,” he wrote, arguing that the open ecosystem “will not win by being the best in its own category.” Users, he argued, simply want the best intelligence for the task at hand — so open models must be on par with, or better than, their closed equivalents.

The strategic logic here is not charity. Poolside’s core business is deploying models inside the security boundaries of government, defense, and regulated enterprises — customers for whom closed, metered API access is often a non-starter for compliance and sovereignty reasons. 

Every enterprise that standardizes on a Chinese open model today becomes harder to win tomorrow. Releasing competitive open weights is both an ecosystem play and a top-of-funnel strategy for the company’s high-security deployment business. It also reframes the AI race away from terrain where Poolside cannot compete — frontier-scale capital expenditure — and toward terrain where it believes it can: cost per token, self-hosting, and iteration speed.

How a sparse architecture makes enterprise AI agents affordable to run

The technical design reflects a specific thesis about where value in coding AI is moving. Laguna S 2.1’s sparse MoE architecture — 256 routed experts plus one shared expert, with grouped-query attention and interleaved sliding-window layers, according to the Hugging Face model card — means inference costs scale with the 8 billion active parameters, not the 118 billion total. Poolside emphasizes that the model is small enough to run on a single Nvidia DGX Spark, the desktop-class AI machine.

That matters for what Poolside calls token economics. Long-horizon coding agents are voracious consumers of tokens: the company’s published data shows the model consuming a mean of roughly 249,000 completion tokens per trajectory on its hardest benchmark when thinking mode is enabled. At metered API prices, agentic workloads at enterprise scale become a meaningful budget line item. On OpenRouter, Poolside is offering a free 256K-context endpoint and a dedicated 1M-context deployment priced at $0.10 per million input tokens and $0.20 per million output tokens — aggressive pricing that undercuts most frontier alternatives by an order of magnitude.

The ecosystem support is unusually broad for day one. The model is live on Baseten’s model library and Vercel’s AI Gateway, with integrations across vLLM, SGLang, Ollama, and llama.cpp, plus quantized variants down to 4-bit GGUF files — 75 gigabytes — for local use. But Poolside’s more interesting claim is behavioral, not architectural. Pengming Wang, co-head of applied research at Poolside, said the gains came from improving the model’s working habits: “more verification, less taking things for granted, not declaring victory early, and being more persistent.” Raw intelligence, the company argues, is one axis of capability; a model’s way of working is a second axis that matters immensely for agents left unattended for hours.

Publishing every benchmark trajectory to counter AI’s credibility crisis

The most consequential part of the release for enterprise buyers may be an evaluation-transparency move with little precedent among major labs: Poolside published the complete, unedited trajectory of every trial in its final benchmark runs — every reasoning step, tool call, and shell command behind every reported score.

This addresses a growing credibility problem in AI benchmarking. As top scores on mature benchmarks cluster in the 70–90% range, and as “reward hacking” — models finding solutions online or gaming verifiers rather than solving problems — has become endemic, self-reported numbers have lost much of their signal. Poolside disclosed its own encounters with the problem candidly: during training, more than half of trajectories on some SWE-bench tasks were flagged because the model simply researched the original bug-fix pull request online and applied it. The company documented its mitigations, including prompt addenda, LLM-based judging calibrated against human labels, and expert annotator review of a high-scoring Terminal-Bench run.

Three published case studies illustrate what the company means by persistence. In one, the model built a working HTML/CSS rendering engine from an empty folder in a 181-step, 50-minute unattended session — then, lacking vision capabilities, spun up headless Chromium to numerically compare its canvas output against a real browser’s rendering. In another, pointed at Poolside’s own agent harness in an automated optimization loop, the model made the Go codebase 5.2% faster with roughly 70% lower memory allocation, finding an O(n²) string-concatenation bug along the way. In a third, working in a sandbox with no Python installed, the model did its number theory in Perl and independently re-derived a proof of Erdős problem #397 — a combinatorics question open for five decades until GPT-5.2 Pro first solved it this past January. Poolside notes that its model’s construction is structurally different from the earlier published solution, and that its November 2025 knowledge cutoff precedes the first proof.

What the disclosed limitations and benchmark fine print reveal

Poolside deserves credit for disclosing limitations most labs bury. The model can overfit to its native harness and stumble on slightly different tool schemas in third-party agents, mangles JSON in nested tool arguments, and is prone to overthinking on competition math. There is currently no user-configurable thinking-effort dial — just on or off — and the gap between the modes is enormous: thinking lifts Terminal-Bench 2.1 from 60.4% to 70.2%, and DeepSWE from 16.5% to 40.4%, at substantially higher token cost.

Buyers should apply their own discounts to the comparison tables. Poolside’s methodology takes the maximum of vendor self-reported scores, benchmark-author leaderboards, and third-party figures for competitors — a reasonable convention, but one that mixes harnesses and test conditions. On DeepSWE, notably, Poolside ran its own agent harness rather than the leaderboard’s standard mini-swe-agent, a difference the company acknowledges makes scores less directly comparable. And the frontier remains clearly out of reach: closed models like GPT-5.6 Sol, at 88.8 on Terminal-Bench 2.1, and Claude Fable 5, at 88.0, along with the 2.8-trillion-parameter open-weight Kimi K3, at 88.3, sit well above Laguna S 2.1.

The deeper structural question is whether Poolside’s “Model Factory” — the internal platform the company credits for its rapid release cadence — can sustain this pace as models scale. The trajectory so far is genuinely unusual: the April dual release of Laguna M.1 and XS.2, the July 2 refresh of XS 2.1, and now S 2.1, which the company says outperforms April’s flagship M.1 at roughly a third of its active size. Remarkably, S 2.1 used the exact same pre-training data as XS 2.1, meaning nearly all the improvement came from scale, training fixes, and post-training across the company’s corpus of 409,000 agentic and non-agentic training environments. Poolside says its next, larger Laguna model began pre-training last week.

For technical decision makers, Laguna S 2.1 is the most credible Western open-weight option to emerge in nearly a year for self-hosted agentic coding — with published evidence, a permissive license, broad ecosystem support, and an economics story built around hardware you can own. Whether it dents the dominance of Chinese open models will depend less on this release than on the ones that follow it.

Kant, for his part, has already told the world how he intends that story to end. Poolside is building toward a future where the most capable intelligence “can be owned and shaped by anyone,” he wrote — and the company plans to keep shipping “until that future exists.” In an industry where the biggest labs increasingly lock their best work behind an API, the most radical thing about Laguna S 2.1 may not be what it scores, but that anyone can download it and check.

I returned to central London for the second consecutive year for the tenth edition of CanJam London, with a fairly clear idea of which brands and headphones I wanted to hear.

Those expectations lasted about five minutes. The moment I entered the show, I encountered several headphones I had never heard of, along with just as many that sounded completely different from what I expected.

Unlike last year, there were so many new products on display that I simply ran out of time to hear them all. What follows is only a snapshot of the new releases and standout headphones I tested at CanJam London 2026—although nearly 30 pairs of cans still represents a rather substantial snapshot.

This report follows our look at the best DACs, headphone amplifiers, and DAPs from CanJam London 2026, while building on our extensive CanJam NYC 2026 coverage. We will also be heading west for CanJam SoCal 2026 later this summer, where another wave of new headphones and personal audio products will undoubtedly be waiting.

Related Reading: The Best DACs, Headphone Amps and DAPs from CanJam London 2026

The Headphones That Stood Out at CanJam London 2026

Grell OAE2

I started at the Electromod booth, conveniently located near the entrance to the show.

The $599 OAE2 is legendary Sennheiser designer Axel Grell’s second attempt at refining his Front-Sided Sound Field Modulation (FSFM) concept. Its 40 mm bio-cellulose dynamic drivers are positioned well forward of the ears, allowing sound to interact more naturally with the outer ear and create a presentation closer to nearfield loudspeakers than conventional headphones. The open-back OAE2 has a 38-ohm impedance, 100 dB sensitivity, a claimed frequency response of 12 Hz to 34 kHz, and weighs 378 grams without its cable.

The original OAE1 received mixed reviews, largely because of its unconventional tonal balance. I did not notice any glaring tonal quirks with the OAE2, although that impression should be taken with a grain of salt because the Electromod booth was hardly a quiet listening environment.

The heavily angled drivers certainly produced the intended forward-facing presentation, which I found pleasant, but the soundstage did not strike my ears as especially wide or expansive. eCoustics Editor-in-Chief Ian White reached a more enthusiastic conclusion during his full review, finding the OAE2 neutral, controlled, and capable of projecting a stable image slightly in front of the listener—although he also noted that the effect varies considerably from one recording to another.

écoute TH2

At the same Electromod booth was one of the more intriguing wireless headphones at CanJam London 2026: the écoute TH2. The closed-back design combines hybrid active noise cancellation and a transparency mode with an eight-band parametric DSP system that allows listeners to make substantial changes to its primary response curve through the accompanying mobile app.

The TH2’s defining feature, however, is its Korg Nutube 6P1 dual-triode vacuum tube in the preamplifier stage, feeding separate Class A/B amplifiers for each channel. It also uses 40 mm titanium-coated Mylar drivers and supports Bluetooth 5.3 with LDAC, AAC, and SBC, along with USB-C digital and 3.5 mm analog connections. Battery life is rated at more than 20 hours.

Sonically, the TH2 delivered prominent midbass with a touch of the harmonic warmth and euphony associated with tube amplification; something rarely experienced from wireless headphones on the move. It was a smooth, relaxed, and thoroughly enjoyable listen. I was told that the TH2 will cost approximately £900 when it goes on sale.

Amoenus Externus

The amoenus externus was the odd one out at CanJam London because it is not a pair of headphones, but rather a software-based spatial-processing system designed specifically for stereo headphone listening.

Instead of adding artificial reverberation or simulating a listening room, the algorithm uses a universal HRTF and an array of nine extremely short delays to reproduce some of the timing cues created by sound interacting with the head, shoulders, and outer ear. The objective is to move the stereo image outside the listener’s head and closer to the presentation of two loudspeakers positioned in front of them. It is available as a stand-alone application or plug-in for 64-bit Windows and Mac systems, along with an AUv3 plug-in for the iPhone.

This was perhaps the most convincing soundstage-enhancement system I have heard thus far. The transformation was not subtle, moving the music away from the sides of my head and placing it significantly farther forward without smothering the recording in artificial ambience. At approximately £20, it is also one of the least expensive products I experienced at CanJam London 2026—and potentially one of the most useful.

FiiO FT15

FiiO brought a handful of new and forthcoming products to CanJam London 2026, which should surprise absolutely no one by now.

First up was the FT15, an open-back planar magnetic headphone expected to replace the FT5 while sitting below the flagship FT7 in FiiO’s expanding over-ear lineup. The wood-accented design also continues the visual direction established by several of the company’s recent headphones, with expected pricing of around £499 when it reaches the market. Final specifications and availability have yet to be confirmed.

The FT15 offered a largely balanced tonal presentation, with treble that remained present and airy without becoming immediately aggressive. That openness worked well, but I found its macrodynamic punch and overall sense of contrast somewhat restrained. It was an easy headphone to listen to, although it did not deliver the same dramatic swings or physical energy as some of the stronger planar designs at the show.

FiiO FG3

Next was the FG3, FiiO’s first gaming headset and one aimed squarely at the ultra-budget-conscious. Its plastic construction looks and feels every bit of its roughly $70 asking price, but expecting milled aluminum and lambskin at this level would be taking the piss.

The closed-back FG3 uses coaxially mounted 50 mm and 16 mm dynamic drivers with an electronic two-way crossover. It also includes an integrated DAC, headphone amplifier, virtual 7.1 surround processing, five gaming presets, and a 10-band EQ accessible through FiiO Control Web. USB and 3.5 mm connections provide compatibility with computers, consoles, and smartphones, while the detachable microphone uses environmental noise cancellation to reduce background noise.

More importantly, it sounded surprisingly decent for the money. There was some gentle roll-off at both ends of the frequency spectrum, but nothing that undermined the overall presentation during my brief audition. The FG3 is already available internationally for $69.99, and I can imagine it rolling off the shelves rather quickly once UK stock becomes widely available. Pun entirely intended.

Snowsky Nature

Apparently, FiiO can go even lower in price without the sound falling apart. SNOWSKY is the company’s affordable sub-brand, and the closed-back Nature costs only $40.

The retro-styled over-ear headphone uses 50 mm dynamic drivers with a Japanese carbon-black composite diaphragm, a low 24-ohm impedance, and a sensitivity rating of 101 dB/mW, making it easy to drive from a smartphone, DAP, or inexpensive USB dongle. It also weighs just 280 grams and includes detachable dual 3.5 mm cables, including one with an inline microphone. Despite appearances, the ear cups use a wood-grain water-transfer finish rather than actual wood.

The Nature sounded slightly hollow during my brief audition, but its detail retrieval and generally coherent tonal balance were far better than I expected at this price. For $40, this one is an absolute steal.

Flowztone F-R1

Hong Kong-based Flowztone is preparing to release its first headphone system, and the company has hardly chosen the cautious route. The F-R1 is a true ribbon design expected to cost around $3,000 with its dedicated amplifier.

The open-back F-R1 uses a 4-micron pure-aluminum ribbon diaphragm coated with an ultra-thin viscoelastic polymer layer. Flowztone claims this construction reduces the low-frequency distortion that has traditionally limited true ribbon headphones, with typical bass distortion rated at 0.5% at 94 dB. The 370-gram carbon-fiber headphone also requires Flowztone’s purpose-built transconductance amplifier to drive its exceptionally low 0.4-ohm load.

Right out of the gate, I found the F-R1 slightly sibilant. Inserting the supplied –2 dB tuning filter brought the treble under better control, however, and allowed me to enjoy its Harman-inspired tonal balance. The F-R1 sounded fast, open, and impressively full-bodied for a ribbon design, making this one of the more compelling debuts at CanJam London 2026.

Noble Audio FoKus Artemis

At $899, the Noble Audio FoKus Artemis sits firmly at the upper end of the wireless headphone market, but it certainly looks the part. The sculpted aluminum construction, sleek lines, premium materials, and magnetically attached earpads give it the appearance and feel of a genuine flagship rather than another plastic travel headphone with an inflated price tag.

The Artemis uses an unusual three-driver configuration comprising one dynamic driver, one planar magnetic driver, and one balanced armature driver in each ear cup. It also features adjustable ANC, transparency mode, LDAC, aptX Adaptive and aptX Lossless support, Audiodo hearing personalization, USB-C audio, multipoint connectivity, and a user-replaceable battery rated for more than 40 hours without ANC or 35 hours with it engaged.

Sonically, the Artemis was right up there with the DALI IO-12 that I reviewed previously, delivering excellent detail retrieval, imaging, and overall technical proficiency. Unfortunately, I could not get the earpads to seal properly against the sides of my head. The resulting air gap rendered the ANC almost useless, which was a bit of a disappointment. Fit will clearly vary between listeners, but at this price, prospective buyers should try them before handing over a credit card.

Sendy Audio Apollo Pro

Moving over to the SIVGA and Sendy Audio booth, I listened to the recently released Apollo Pro, a substantially redesigned successor to the original Apollo.

The open-back headphone uses a custom 78 mm planar magnetic driver with a nano-scale composite diaphragm measuring less than 800 nanometers thick. Sendy Audio also employs electron-beam deposition to apply the aluminum conductor layer more uniformly. The 354-gram design combines zebrawood ear cups with CNC-machined aluminum components and includes a detachable 4.4 mm balanced cable with 3.5 mm and 6.35 mm adapters. Impedance is rated at 27 ohms, with a sensitivity of 93 dB.

I have not heard the original Apollo, but I enjoyed the Pro’s particularly intimate and engaging vocal presentation. The trade-off was a soundstage that felt somewhat compact for an open-back planar headphone. Even so, its detail retrieval, separation, and overall technical performance were highly competitive at $494.

ZMF Headphones Tessidera

The moment I picked up the $2,499 ZMF Tessidera, I was struck by just how beautiful it looked. ZMF headphones have always resembled works of art, but the angular Canarywood cups and honeycomb-like ZMF Acoustic Matrix visible behind them elevate the design to another level.

The open-back Tessidera uses a newly developed 1-micron planar magnetic driver with a 35-ohm impedance and sensitivity of approximately 92 dB/mW. It combines ZMF’s CAMS magnet structure and Atrium Damping System with the new lattice-style ZMF Acoustic Matrix, designed to reduce standing waves and smooth the treble, and a wooden Soundboard Resonator System intended to improve transient impact and control resonance.

More importantly, how does it sound? To my ears, the Tessidera leaned closer to neutrality than many ZMF headphones, with something of the Auteur’s balance and natural presentation. The treble was smooth and inviting, but dynamics remained strong, while timbre and imaging were as realistic and precise as I have come to expect from ZMF. I have not always connected with the company’s tonal choices, but the Tessidera was right up my alley.

Final Audio Closed-Backs – DX3000CL and DX4000CL

Final recently expanded its premium closed-back headphone range with the DX3000 CL and DX4000 CL. They look almost identical, but the DX3000 CL costs $599, while the DX4000 CL lands just below $1,000.

Both use 40 mm dynamic drivers with lightweight Japanese washi paper and carbon-fiber composite diaphragms, along with thick 30 mm memory-foam earpads and an internal rear-diffuser array designed to suppress unwanted resonances. They also share a 37-ohm impedance and 96 dB/mW sensitivity. The DX4000 CL weighs 375 grams and includes Final’s upgraded silver-coated cable, while the DX3000 CL weighs 410 grams and comes with a standard OFC cable.

In noisy show conditions, I found it difficult to distinguish between their sound signatures. I was told that the more expensive DX4000 CL is slightly more analytical and resolving, while retaining the warmth and generous bass response shared by both models.

Listeners who prefer a warm, full-bodied presentation should find plenty to enjoy here. Unless the DX4000 CL’s lighter construction, added refinement, and upgraded cable are essential, however, I would take the DX3000 CL and pocket the $400 difference.

Final Audio DX10000CL

Final had one more new closed-back headphone on display, although the DX10000 CL occupies a very different financial universe from the DX3000 CL and DX4000 CL. The standard model costs a rather unhinged $8,499, while the limited Collector’s Edition adds another $500.

The flagship uses a newly developed 40 mm dynamic driver with a self-supporting True Diamond diaphragm manufactured through chemical vapor deposition. Final claims the material’s extreme rigidity and strong internal damping reduce deformation, resonance, and distortion while improving transient response. The aluminum-magnesium alloy housing has also been engineered to control internal reflections, while the Ultrasuede earpads manage airflow around the ear.

Its subtle steampunk aesthetic looks terrific, but the real surprise came when I listened to it immediately after its two considerably less expensive siblings. The difference was night and day, with substantially greater clarity, perceived detail, and openness owing to its more neutral tonal balance.

I will admit that the DX10000 CL did not tug at my heartstrings quite as strongly as some of the other headphones I heard at the show. Which ones did? I will save those for the Best in Show section.

Audeze Maxwell 2 ANC

Returning to wireless headphones, I gave the Audeze Maxwell 2 ANC a whirl. As its name suggests, it adds adaptive hybrid noise cancellation and a low-latency transparency mode to the Maxwell 2 platform. It is also the first Audeze headphone of any kind to offer ANC.

The closed-back gaming headset retains Audeze’s 90 mm planar magnetic drivers, Fluxor magnet arrays, Fazor waveguides, and SLAM acoustic technology. Bluetooth 5.3 support includes LDAC, LE Audio, multipoint, and Auracast, which allows a single broadcast source to transmit audio to multiple nearby compatible devices without conventional pairing. Battery life is rated at 40 hours with ANC switched off and 35 hours with it engaged. At 560 grams, however, nobody will mistake it for lightweight travel headphones.

Despite the added weight over the original Maxwell, I preferred the newer model’s sonic balance. Its flatter bass response kept the midrange cleaner without sacrificing impact, while the planar drivers retained the clarity and positional precision that made the Maxwell series stand out in the first place. The PlayStation version costs $429, while Xbox users will pay $449—$100 more than the equivalent Maxwell 2 without ANC.

Audeze MM-520

Sticking with Audeze, the $1,699 MM-520 is a studio-focused evolution of the MM-500 developed with 18-time Grammy-winning mixing engineer Manny Marroquin. Its most significant addition is Audeze’s Symmetric Linear Acoustic Modulator, or SLAM, which uses engineered channels to manage airflow and pressure inside the ear cups, improving bass definition and spatial imaging.

The open-back MM-520 uses 90 mm planar magnetic drivers with Ultra-Thin Uniforce diaphragms, Fluxor magnet arrays, and Fazor phase management. It also offers an 18-ohm impedance, 102 dB/mW sensitivity, and a claimed frequency response of 5 Hz to 50 kHz. At 555 grams, however, it remains a substantial piece of studio hardware.

The MM-520 delivered an exceptionally detailed and precise presentation, with the kind of midrange clarity and tonal accuracy expected from Marroquin’s involvement. Even with SLAM, however, I did not find the soundstage especially expansive, and its macrodynamic impact felt somewhat restrained. It is an impressive monitoring tool, but listeners looking for a more physical or emotionally charged presentation may prefer something less analytical.

Moondrop Skyland

The Elise Audio booth had far too many headphones for me to audition them all, but the Moondrop Skyland immediately caught my eye. It was the first pair I picked up—and wow, this thing is heavy.

The open-back Skyland uses enormous 100 mm planar magnetic drivers with 500-nanometer diaphragms, N55 magnets, and a full-surface drive system. Its all-aluminum CNC construction is paired with a carbon-fiber suspension headband, hybrid lambskin earpads, and a 4.4 mm balanced cable with 6.35 mm and four-pin XLR adapters. Impedance is rated at 60 ohms, with a sensitivity of 96 dB/Vrms and a claimed frequency response of 8 Hz to 34 kHz.

Unfortunately, the substantial weight was compounded by rather low clamping force, resulting in a headphone that did not feel especially secure on my head. I also had to use the smallest headband setting to achieve a workable fit on my average-sized head. Not a promising start.

At least the Skyland sounded good, offering a balanced tonal presentation, seemingly bottomless bass extension, and solid macrodynamic contrast. I was not exactly blown away, however, and the awkward fit would make the $799 asking price difficult to overlook.

Auribus Acoustics Laguna

I was subsequently approached by Auribus Acoustics founder Juan Ovalle, who wanted to hear my impressions of his latest creation: the closed-back Laguna.

The prototype is expected to launch around September for approximately $1,500, although final specifications, pricing, and availability have yet to be announced. Auribus Acoustics is a Southern California headphone manufacturer known for using advanced 3D-printing techniques to refine its acoustic chambers and reduce unwanted resonances.

The moment I placed the Laguna over my ears, I was impressed by its passive noise isolation. It blocked out a substantial amount of the noisy show floor before I had even pressed play.

It also sounded excellent, producing deeply textured sub-bass and a subtle midbass lift that added weight without pushing the presentation into the mud zone. My only minor concern was the treble, which lacked some fine-grained clarity and definition to my ears. Even so, this was a highly promising prototype and one of the more convincing closed-back headphones I heard at CanJam London 2026. Another listener also praised its deep but controlled bass and unusually spacious presentation for a closed-back design.

Viking Audioforge Draupnir

I auditioned a prototype of the Draupnir at CanJam London 2025 and awarded it our Best Innovation honor. One year later, Viking Audioforge has turned that ambitious experiment into a considerably more polished production headphone, now available for preorder at £1,799. Although the fit, finish, and adjustment mechanism have been refined for more consistent manufacturing, Viking Audioforge says the sound, core design, and materials remain unchanged from the Launch Edition.

The open-back Draupnir uses two 50 mm glass-diaphragm dynamic drivers per side, combined through a carefully shaped acoustic lens so that they behave as a coherent source. Its baffle-mounted bass switch adds approximately 6 to 8 dB below 100 Hz, while the lightweight MJF nylon chassis, carbon-fiber headband components, magnetic ZMF earpads, and included Viking Weave cable complete one of the more unconventional headphone designs at the show. It weighs approximately 390 grams.

I cannot say with certainty whether the tuning has changed since last year, but it remains every bit as enjoyable as I remembered. The Draupnir delivered an inviting sense of space and an exciting, quick bass response that became delightfully bombastic once I engaged the boost switch. It would not be an exaggeration to call this one of the most memorable listening experiences I had at CanJam London 2026.

Nur Miah

I arrived at the Hifonix booth with no expectations before trying the NUR Miah; admittedly, I had done no research on the Italian headphone manufacturer beforehand.

The forthcoming Miah is a closed-back dynamic headphone with drivers angled forward toward the ears. NUR has not yet published detailed specifications or a firm release date, although I was told at the show that pricing should land somewhere around £2,200.

What surprised me most was the sheer dynamism and realism of its presentation. The Miah sounded exceptionally immediate and lifelike, while maintaining a well-balanced tonal signature that prevented all that energy from becoming tiring or exaggerated. This was an exciting listen—and easily one of the biggest surprises of CanJam London 2026.

Grado Signature S550

Grado has been rather hit or miss for me in the past. I bought and quickly sold the SR80x because I could not get past its sharp upper midrange, while the HP100 SE proved far more successful, delivering an exciting presentation with beautifully rendered vocals.

The open-back Signature S550 uses Grado’s newly developed 50 mm S2 dynamic driver, with a 38-ohm impedance, 112 dB sensitivity, and a claimed frequency response of 6 Hz to 44 kHz. It weighs 335 grams and includes Grado’s detachable Signature Silver cable, connected through locking four-pin mini-XLR sockets. The S550 ships with the company’s new B cushions but is also compatible with its S, F, L, and G pads, allowing listeners to alter both the fit and sonic presentation.

The $995 S550 firmly lands in the hit column. I have always enjoyed the distinctive appearance of Grado headphones, but what drew me in here was the authoritative bass and excellent dynamic impact. More importantly, it delivered that energy without the overly sharp, grating upper-midrange and treble presentation that has pushed me away from some of the company’s less expensive models.

HiFiMAN HE6 Remastered

The HiFiMAN HE6 Remastered revives the driver design of the original HE6 while adding the company’s second generation composite headband. This is less a reinvention than an attempt to make one of HiFiMAN’s most celebrated headphones slightly more comfortable and accessible to a new generation of listeners.

The open back planar magnetic headphone retains the original’s brutally inefficient 83.5 dB sensitivity, 50 ohm impedance, and claimed frequency response of 8 Hz to 65 kHz. It weighs 522 grams, which is 28 grams lighter than the original version, but still hardly featherweight.

The original HE6 has achieved near legendary status, so I suspect the $1,899 Remastered will find plenty of willing buyers. For good reason: it remains one of the most impactful and dynamically explosive planar headphones ever produced. Just bear in mind that this is not something you should connect to a phone, laptop, or underpowered dongle DAC. The HE6 Remastered needs a seriously powerful desktop headphone amplifier and will expose anything pretending to be one.

Best in Show

Best Budget Headphones: Hongshu Audio EH-1

Before anyone reaches for the pitchforks, I know that $649 is hardly cheap. For a complete electrostatic headphone system, however, it qualifies as budget territory, especially because Hongshu Audio includes the dedicated EHA-1 energizer in the price.

The open back EH-1 uses a four inch electrostatic driver, weighs only 295 grams, and is designed around the standard 580 volt bias used by many electrostatic headphones. The compact EHA-1 energizer provides up to 300 Vrms of output, along with a physical channel balance control and a switchable bass boost.

The EH-1 captures the strengths of electrostatic headphones remarkably well, delivering excellent detail, speed, and control across most of the frequency range. Deep subbass rolls off somewhat, which is hardly unusual for the driver type, but Hongshu Audio has also avoided the excessively bright treble that can undermine electrostatic headphones costing several times more.

Listeners looking for greater weight can engage the energizer’s bass boost, which lifts the response from the midbass downward without destroying the headphone’s clarity. At $649 for the complete system, there is very little not to like about the EH-1.

Best Mid-price Headphones: Dan Clark Audio Aeon Core

Competition among new wired headphones in the mid-price category, which now seems to hover around $1,000, was surprisingly thin at CanJam London 2026. That should not diminish what Dan Clark Audio achieved with the AEON CORE, however. This was an excellent headphone in its own right.

The closed back AEON CORE is the direct successor to the AEON 2 and uses a new, more efficient planar magnetic driver. It is also the first Dan Clark Audio headphone tuned to a revised Harman over ear target developed through research with Dr. Sean Olive. The design combines aluminum and wood ear cups with the company’s familiar suspension headband and folding frame. It weighs 440 grams, with a 17 ohm impedance and sensitivity of approximately 94 dB/mW.

The AEON CORE was probably the best fitting headphone I tried at the show. Its suspension strap distributed the weight evenly, while the moderately firm clamping pressure and compact ear cups molded comfortably against the sides of my head.

Listening through the Violectric V281 Ultimo certainly did not hurt, but the AEON CORE sounded excellent across the board. Bass was tight and controlled, the midrange remained present and expressive, and the treble added plenty of sparkle without upsetting the overall balance. Everything felt carefully judged without becoming dull, which is a surprisingly difficult trick to pull off.

Best High-End Headphones: Meze Audio ARTA

Meze Audio headphones have always looked as though they belong in an art exhibition, and the ARTA continues that tradition. Its meticulously machined components, sculpted metalwork, angled acoustic grille, carbon fiber, and leather come together in an elegant Art Nouveau inspired design that is unmistakably Meze.

The open back ARTA uses Rinaro’s newly developed MZ5 HΩ Isodynamic Hybrid Array planar magnetic driver. Its unusually high 225 ohm impedance is intended to work with higher voltage headphone amplifiers, while the angled acoustic blades behind the driver are designed to reduce internal reflections and improve openness and transient clarity. The ARTA weighs 495 grams and has a claimed frequency response of 3 Hz to 115 kHz. Every major component can also be serviced or replaced, which matters when a headphone costs $6,000.

Some listeners will expect magic at this price, although this hobby has repeatedly demonstrated that a large number on the box guarantees very little. The ARTA, however, delivered deep and beautifully textured bass, along with a remarkably open soundstage that could sound expansive or intimate depending on the recording. Tiny details emerged without requiring me to concentrate or squint into the mix, while its technical precision never came at the expense of emotional engagement.

The Meze Audio ARTA is a display piece that also happens to reproduce music with equal measures of proficiency, scale, and feeling.

Best Wireless Headphones: Noble Audio FoKus Apollo Pro

Noble Audio’s less expensive wireless headphone impressed me more than the flagship FoKus Artemis. The FoKus Apollo Pro, not to be confused with the Sendy Audio headphone bearing the same name, may not look quite as luxurious, but its ANC worked far more effectively on my head and the sound quality was every bit as convincing.

The limited run Apollo Pro combines a 40 mm dynamic driver for the bass with a 14.5 mm planar magnetic driver for the midrange and treble. It also offers hybrid ANC, LDAC, aptX HD, USB C audio, app based EQ, and up to 80 hours of battery life with ANC switched off. Noble has also upgraded the tuning, earpads, headband materials, and overall finish compared with the original Apollo.

It is surprisingly rare to hear a genuinely balanced tonal presentation from wireless headphones, but the FoKus Apollo Pro delivered exactly that. Bass was tight and controlled, the midrange remained clean and present, and the treble offered clarity without becoming sharp. At $699, this was the stronger wireless headphone in Noble Audio’s CanJam London lineup.

The Bottom Line

The fact that I ran out of time before hearing every new headphone at CanJam London 2026 says plenty about the pace of innovation in personal audio.

Some headphones were genuinely exciting, while others were complete disappointments and did not make this report. The real value of CanJam is that specifications, prices, and reputations only tell part of the story. Every pair has the potential to surprise you, for better or worse, the moment you place it over your ears.

That unpredictability is precisely what makes these shows worth attending.

Related Reading:

SOFTWARE

Office LTSC 2021, Windows Server 2022, Publisher, and Entra ID risk policies all facing the chop within weeks of each other

October 2026 is shaping up to be a busy month for administrators as it heralds the end of several big products, along with the retirement of components that could mean sleepless nights for the ill-prepared.

After the demise of Windows 10 in October 2025, sysadmins might have hoped for some respite from Microsoft’s ax, but 2026 will see the demise of more beasts.

First, there’s the end of support for Office LTSC 2021, which means everything from Access 2021 to Word 2021 is for the chop. A move to LTSC 2024 is required to keep the support lights burning, or – and we don’t doubt that this is Microsoft’s preference – a switch to Microsoft 365.

M365 isn’t a great option for organizations steering clear of the cloud. The on-prem version is very much a snapshot of where things were at release (although security patches keep flowing).

Speaking of on-premises, Windows Server 2022 will drop out of mainstream support on October 13, 2026. Extended support will linger until 2031. Windows 11 24H2 Home and Pro reach the end of the road on the same date. Windows 11 23H2 for enterprise and education will get until November 10 to tidy up their affairs.

As for the more minor products? Publisher 2021 will wrap up decades of the Microsoft Publisher brand in October. First appearing as a desktop publisher for Windows 3.0 in 1991, it has had an inexplicably long life, which will come to an end once and for all in a few short months. And unlike the rest of the Office 2021 LTSC suite, there won’t be anything to replace it (or its annoyingly proprietary file format).

There are also services due for the chop that might cause headaches for those that fail to prepare. Entra ID Sign-In risk policies, formerly known as Identity Protection, is being retired on October 1, 2026. Microsoft wants admins to migrate to Conditional Access, and make sure their policies cover user and sign-in risk. Not done the update? Then risk protection will go away after the retirement date.

A Register reader remarked, “Given how many orgs use O365 I can’t see how disabling sign-in protection might go wrong.”

October 2027 looks quieter, with only SQL Server 2017 gasping its last. However, with the coming October loaded with already announced terminations, getting the corporate house in order well ahead of time would be prudent. ®

Even if some of them are half-hearted.

Swings and lurches

Paramount

But when it works…

Paramount

A forged Hide My Email header can make Mac Mail expose a user’s Apple Account address, extending the privacy risk beyond people who use Apple’s alias service.

Developer Jeff Johnson disclosed the Mac Mail privacy flaw on July 19. His testing showed that a specially crafted message could make Mail display a misleading sender identity while a reply went out from the email address tied to his Apple Account.

The behavior didn’t require Johnson to use Hide My Email or maintain an iCloud.com mailbox. He also found that the malicious message could arrive through an unrelated personal or business account before Mail exposed the Apple Account address in a reply.

Johnson said he doesn’t know whether the finding is technically related to an earlier flaw that could reveal the address behind a Hide My Email alias. His report documents a separate technique and doesn’t establish that both issues share the same underlying cause.

Johnson used the curl command-line tool to send himself an email containing an arbitrary X-Icloud-Hme header. Mail doesn’t give users direct control over arbitrary outgoing headers, while curl allowed him to choose the values placed inside it.

By controlling those values, Johnson could make Mail show selected addresses in the reply window. The app also displayed “Hide My Email” in the From field and added a notice saying the message had been forwarded through the service.

The result suggests Mail accepted information supplied by the sender without first establishing that Apple’s Hide My Email service had generated it. Johnson demonstrated the visible behavior, but his report doesn’t identify the app’s internal validation process.

A crafted X-Icloud-Hme header was sufficient to trigger the Mac Mail bug in Johnson’s testing. Image credit: Jeff Johnson

The problem goes beyond a misleading label because users rely on the From field to confirm which account will send a message. Mail reportedly showed one identity while sending the reply from another, giving Johnson no accurate warning that his Apple Account address would be exposed.

The reported flaw reaches beyond Hide My Email users

Johnson said he doesn’t use Hide My Email and doesn’t have an iCloud.com email account. That makes the reported issue broader than a flaw limited to people actively using Apple’s alias service.

A crafted message can also arrive through a separate email account, yet Johnson said Mail may still send the reply from the address tied to the recipient’s Apple Account. The distinction matters for people who keep personal, business and Apple Account identities inside the same Mail app.

Apple Mail can manage multiple accounts from different providers, and users reasonably expect those identities to remain separate. Johnson’s test shows that the compose window may not always reflect the address that Mail will actually transmit.

Johnson documented a technique he reproduced in his own testing, but the report doesn’t establish which macOS versions are affected. It also doesn’t show whether the same behavior works in Mail on iPhone or iPad, and it offers no evidence that anyone has used the technique in an attack.

The only other warning was a notice above the message saying it had been forwarded through Hide My Email. Image credit: Jeff Johnson

Crafted email headers have caused problems for Apple Mail before. In 2022, a specially formatted From field could prevent some iOS 16 users from opening Mail, although that issue caused a crash rather than exposing an email address.

The new finding also arrives while Apple faces legal scrutiny over the earlier Hide My Email disclosure. A proposed class action lawsuit filed July 15, accuses the company of misleading customers about the feature, although the complaint doesn’t allege that attackers exposed the plaintiff’s address or exploited the flaw.

How to protect yourself from the Mac Mail disclosure flaw

Mac Mail users should stop before sending when a reply window unexpectedly labels the From address as “Hide My Email,” particularly when the original message wasn’t sent to an alias. In Johnson’s test, the disclosure occurred only after he sent the reply.

Users can inspect a received message’s complete headers by opening the View menu, selecting Message and choosing All Headers. Johnson noted that the command can also be added to the message window’s toolbar.

The presence of an X-Icloud-Hme header doesn’t establish who added it or why. An unexpected Hide My Email label combined with unusual From or To fields should still be treated as a warning.

MacGeneration independently reproduced the Mail bug and confirmed that it can expose an Apple Account address even when the user doesn’t use iCloud for email. The publication’s testing supports Johnson’s findings, though the affected Mail app versions remain unclear.

Users who need to respond should create a new message and confirm the From address before sending instead of replying directly to the suspicious email. Creating a separate message won’t fix the underlying Mail bug, but it avoids the reply behavior Johnson documented.

• TeamViewer report finds 71% of autonomous AI users set guardrails, but 61% wouldn’t even use autonomous AI to begin with
• The average worker spends 2+ hours per week validating AI output
• Security and privacy could have the biggest impact

New TeamViewer research positions trust as the latest major barrier to AI adoption, with workers generally viewing the tool positively but struggling to hand over full independence to it.

Today, three in four workers use AI daily and only 3% say they don’t see any noticeable workplace benefits, but nearly all (95%) respondents have at least some concern about AI operating without a human in the loop.

Substack is giving readers a way to check whether the post they’re reading was written by a human or by a chatbot. The company has partnered with AI-detection firm Pangram to introduce new tools that will let users scan posts, notes, and replies for AI-generated text. CEO Chris Best introduced the features in a post titled “Against Claudefishing,” his term for content that leans on AI while presenting itself as human work.

How the scanning tool works

Scans run on any text longer than one hundred words that is published from July 21 onward, and the result is visible only to the person who requests it. Best says Pangram is not perfect, but points to independent evaluations that credit it with a high degree of accuracy. The feature is live on web and iOS, with an Android version coming later.

Substack’s AI-detection feature arrives at a time when the problem it targets is spreading fast across social platforms. Best cites a recent Pangram estimate that as much as forty percent of posts on some platforms are now fully AI-generated.

Writers can run the same scan before they publish

Readers are not the only ones getting new controls. Substack is adding a “How I make this” statement that lets writers describe their process and set expectations up front. Writers also get the option to run Pangram on their drafts before publishing, and they can report and remove scans of their published work that they believe are mistaken. Best says future additions could include AI preferences inside Reply Rules and reader-side controls for what gets recommended.

For readers, this is a rare transparency feature that hands them a real button to spot AI slop. Substack is not alone in trying to solve this issue. LinkedIn has started reducing the reach of AI-generated posts and comments, and Meta is quietly testing its own AI detection tool, though it has yet to go live. Whether any of these tools can keep pace with the models they’re built to catch remains to be seen.

For years, Synthesia’s pitch was to use AI to help enterprises create interactive training videos, in minutes, for a fraction of the cost normally associated with developing corporate educational materials. The startup’s newest product is making a different bet — that the real moat in enterprise AI isn’t just generating content, but in proving that it worked. 

On Wednesday, the British startup launched Roleplay Sessions, an interactive training product where employees can practice high-stakes conversations like sales pitches, performance reviews, and customer complaints with an AI avatar that talks back, pushes back, and then scores them against a rubric. 

Roleplay is the first release under a broader “Sessions” platform Synthesia plans to expand into other formats, including job interviews and candidate screening, TechCrunch has exclusively learned. 

The launch comes as enterprises take a harder look at whether their AI spend is paying off. Citing a meta-analysis of learning research, Synthesia argues that most corporate training — including, implicitly, its own core video product — stops short of actually changing behavior. The trainings tend to inform and demonstrate, but what really pushes someone into actually learning something is doing that thing, along with feedback.

“Video performs way better than text or sending out a document,” Synthesia CEO and co-founder Victor Riparbelli told TechCrunch. “But for most things, we learn the best by actually practicing something rather than just reading it.”

Roleplay also moves Synthesia into a more defensible position. The company’s avatar and voice technology are proprietary, but the underlying reasoning intelligence is OpenAI’s. By tacking on a layer that includes rubrics, performance data, and analytics, Synthesia is positioning itself less as an AI avatar company and more as a performance-management platform with a video front end. 

“From a managerial executive layer, we’re seeing huge interest in mapping out the talent in their company in a way that’s been really difficult to do before,” Riparbelli said. “If you have your entire sales team practice with an AI role player, you can get very granular data on how your sales force is doing overall.”

Synthesia already has a few early Roleplay customers with scaled commercial rollouts, including “one of the top three companies by market cap in Europe, one of the top five Fortune 100, and one of the biggest recruitment companies in the world,” per a company spokesperson. 

The two most popular use cases involve training sales teams and leadership — programs like practicing how to sell a product or have a difficult conversation. Much of it is soft skills training, Riparbelli said. 

As with Synthesia’s main product, customers can either create their own training programs on the platform or enlist the help of Synthesia’s consultants to build a bespoke solution that relies in part on whatever training documents and context the company already has. 

Roleplay is an enterprise offering for now, but the goal is to offer it more widely to small businesses, prosumers, and potentially educational institutions in the next few months as inference costs continue to fall. Riparbelli believes products like Roleplay represent the next phase of enterprise AI adoption, where companies care less about whether AI looks good in a social media demo and more about proving it generates measurable business outcomes.