Plutonic Rainbows

Twelve days after Opus 4.6 landed, Anthropic released Sonnet 4.6 at the same $3/$15 per million tokens as its predecessor. The benchmarks tell the story: 79.6% on SWE-bench Verified against Opus's 80.8%. A gap of 1.2 points. For 60% of the price.

Computer use is where it gets embarrassing for everyone else. Sonnet 4.6 scores 72.5% on OSWorld-Verified. GPT-5.2 manages 38.2%. That's not a competitive gap — that's a different sport.

Early testers preferred Sonnet 4.6 over Opus 4.5 59% of the time. The previous flagship. Beaten by its own cheaper sibling released three months later. The pattern keeps repeating across the industry — the mid-tier closes the gap, the flagship justifies itself for fewer and fewer workloads, and the pricing structure starts to look like a loyalty tax.

I'm writing this on Opus 4.6. I'm not sure why.

Lisa Graham, somewhere against a late-eighties coastline — cobalt blazer, exaggerated lapels, leather gloves, a single corsage holding it all together. Originally the cover of Marie Claire Spain, September 1988 — the kind of editorial that makes you wonder why anyone stopped dressing like this.

Somewhere around the third hour of listening I stopped taking mental notes. That's usually how I know a system is working. The Exposure 2510 integrated amplifier and ATC SCM11 standmount speakers had been running together for a few weeks by that point, and the initial critical ear — the one that listens to the equipment rather than through it — had gone quiet. What remained was just music, delivered with a directness that caught me off guard given the combined price.

The 2510 is a 75-watt Class A/B integrated from Exposure's factory in West Sussex, a company that has been building amplifiers in England since 1974 without ever chasing fashion. The design philosophy is almost aggressively simple: four controls on the front panel (input selector, volume, and two buttons for standby and mute), five line inputs, a built-in MM phono stage, and a pre-amp output if you ever want to add a power amplifier later. No DAC. No streaming. No tone controls. The circuit uses discrete components with high-quality VIMA capacitors in the signal path, and the through-hole assembly is still done by hand in England. It weighs six kilograms, which feels light until you hear what it does with those watts.

The SCM11 is ATC's entry-level standmount, though calling anything ATC makes "entry-level" feels slightly misleading. This is a sealed-box two-way design with a 150mm mid/bass driver using ATC's Constrained Layer Damping technology — a technique that reduces resonance within the cone itself rather than trying to tame it externally — and a 25mm soft-dome tweeter with a neodymium magnet and proprietary alloy waveguide. Sensitivity is 85dB, which means they need decent amplification, and nominal impedance sits at 8 ohms. ATC builds every driver in-house at their Stroud factory in Gloucestershire, winding the voice coils by hand. Both companies are separated by roughly 150 miles of English countryside and share an almost identical reluctance to overcomplicate things.

What strikes me about this pairing is how the warmth of the Exposure meets the neutrality of the ATC without either quality cancelling the other out. The 2510 has a gently rich midrange that could, with the wrong speakers, tip into softness or veil detail. The SCM11 has a forensic transparency that could, with the wrong amplifier, sound lean or clinical. Together they arrive at something I can only describe as honest warmth — the kind of presentation where a piano sounds like wood and hammers and resonating strings rather than a synthetic approximation of those things. Voices sit in the room with real weight. Not projected, not recessed. Just present.

The sealed-box loading of the SCM11 is doing important work here. Ported speakers can produce more low-end extension, but the bass they deliver often trades speed for depth. The SCM11 rolls off below 56Hz, so you're not getting subterranean weight, but what's there — everything from kick drums to upright bass to the low growl of a cello — arrives with a tightness and control that makes ported designs at this price sound flabby by comparison. I spent an evening working through some of my ripped collection via the network player and the sealed enclosure kept up effortlessly across genres. Dense electronic music, sparse acoustic recordings, everything in between. No port chuffing, no overhang. Just grip.

Seventy-five watts into 85dB-sensitive speakers might look marginal on paper, but I never came close to running out of headroom. The 2510's power supply is doing more useful work than the raw number suggests — it has current delivery that belies the specification sheet. I rarely pushed the volume past the halfway mark even in a medium-sized room with the speakers on stands well clear of the rear wall. Forum users who've paired earlier Exposure models with the SCM11 report the same experience: instruments and voices becoming discernible in a way they weren't before, an openness that invites you to keep listening without fatigue.

The midrange deserves its own paragraph. Both components prioritise this region, and it shows. ATC's CLD driver was designed specifically to clean up the critical band between 200Hz and 3kHz where most musical information lives, and the Exposure feeds it with a signal that's detailed without being etched. The result on vocal recordings is striking — not in an audiophile-cliché way where you suddenly hear a singer's saliva, but in the sense that phrasing and dynamics come through intact. You hear the intention behind a vocal performance rather than just the notes. I went back to records I'd dismissed as flat or poorly mastered and found layers of nuance sitting right there in the mix, previously masked by less resolving equipment.

I keep coming back to the value proposition. The 2510 retails for approximately £2,100 and the SCM11 sits around £1,650. Call it £3,750 for the pair before cables and stands. In a market where a single component can easily exceed that figure, what you're getting here is two British manufacturers' distilled engineering philosophy — decades of refinement expressed as restraint rather than feature creep. Neither product tries to be everything. The amplifier amplifies. The speakers convert electrical signal to sound pressure. And the narrow focus pays off in the listening chair.

There are limitations I should acknowledge. The SCM11's bass rolls off where a floorstanding speaker is just getting started, so if your musical diet demands foundation-shaking lows, you'll either need a subwoofer or different speakers. The 2510 offers no digital inputs whatsoever, which means your source needs its own DAC — a consideration I've already wrestled with when assembling this chain. And the ATC's 85dB sensitivity means they're not the right choice for a feeble amplifier tucked inside a sideboard. They want proper amplification and proper placement.

None of that diminishes what this combination achieves within its design envelope. The Exposure 2510 and ATC SCM11 share something I rarely encounter in mid-priced hi-fi: the absolute absence of a weak link. Most systems at this level have one component that's clearly punching above its weight while another holds things back. Here, the engineering intent is so closely aligned that everything arrives at the same standard simultaneously — detail, dynamics, tonal accuracy, spatial coherence. I stopped thinking about the equipment and started thinking about the music, which is the only review metric that actually matters.

Anthropic ran a Super Bowl ad mocking OpenAI for putting ads in ChatGPT. Altman fired back on X calling Anthropic "authoritarian." OpenAI poached an Anthropic safety researcher the same week it shipped a product that looks like a direct response to Claude Code. The whole thing has the energy of two restaurants opening across the street from each other and spending more time reading each other's menus than cooking.

These companies have genuinely different strengths. Anthropic owns 40% of enterprise LLM spend and builds tools developers actually trust with production code. OpenAI has consumer reach nobody else touches and multimodal ambitions that stretch well beyond text. Both positions are defensible. Neither requires obsessing over what the other shipped on Tuesday.

I wrote about speed becoming the only moat last month. The same logic applies to rivalry as a strategy — it narrows your field of vision to exactly one competitor while the market fragments around you. Meta, Google, and a dozen open-source efforts don't care about your Super Bowl ad.

GPT-5.3-Codex helped debug its own training. OpenAI said it plainly: "the first model that was instrumental in creating itself." That was ten days ago. This week, ICLR announced their first workshop dedicated entirely to recursive self-improvement, scheduled for Rio in April. Google's AlphaEvolve already discovered algorithmic improvements that beat Strassen's fifty-six-year-old matrix multiplication record. The pieces are landing on the board faster than anyone expected.

Recursive self-improvement — systems that modify their own code, weights, prompts, or architecture to become more capable, then use that increased capability to improve themselves further — has been a thought experiment for decades. Eliezer Yudkowsky warned about it. Nick Bostrom built philosophical scaffolding around it. And for most of that time it remained comfortably theoretical because the systems weren't good enough at the one thing the loop requires: writing better software than the software that already exists.

That constraint is dissolving. Not because we've achieved some sudden breakthrough in machine consciousness or general reasoning, but because the narrow version of self-improvement turns out to be enough to matter. A model doesn't need to understand itself philosophically to optimise its own training pipeline. It just needs to be good at code. And the current generation is good at code.

The METR data makes the trajectory explicit. AI task-completion horizons have been doubling every four to seven months — depending on which estimate you trust — for the past six years. If that holds for another two years, we're looking at agents that can autonomously execute week-long research projects. Another four years and it's month-long campaigns. The trend line itself isn't the alarming part. The alarming part is that the trend doesn't need to hold perfectly. Even if progress halves, the capability gap closes on a timeline measured in quarters, not decades.

Dean Ball put it starkly in his recent analysis: America's frontier labs have begun automating large fractions of their research operations, and the pace will accelerate through 2026. OpenAI envisions hundreds of thousands of automated research interns within nine months. Dario Amodei cites 400% annual efficiency gains from algorithmic advances alone. These aren't wild extrapolations from startup pitch decks. These are the people running the labs describing what they see happening inside their own buildings.

However. There's a constraint that rarely gets enough attention in the acceleration discourse. Self-improvement only generates reliable gains where outcomes are verifiable. Code that passes tests. Algorithms with measurable performance. Training runs with clear loss curves. The loop works brilliantly in these domains because you can tell whether the modification actually helped. The system generates a change, measures the result, keeps or discards. Simple evolutionary pressure.

The loop breaks — or at least stumbles badly — when it encounters domains where verification is ambiguous. Alignment research. Safety evaluation. Novel hypothesis generation. The things that arguably matter most for whether recursive self-improvement goes well or catastrophically. A system can optimise its own matrix operations all day. Whether it can meaningfully improve its own ability to recognise its blind spots is a much harder question, and I suspect the honest answer is no.

So when will genuine recursive self-improvement arrive? It depends on what you mean. The narrow version — models improving their own infrastructure, training pipelines, and deployment tooling — is already here. GPT-5.3-Codex is doing it in production. The medium version — agents that systematically discover architectural improvements and better training recipes — is probably twelve to eighteen months out, conditional on the METR trendline holding. The strong version — a system that improves its own reasoning capabilities in open-ended domains, including the ability to improve its ability to improve — remains genuinely unclear. I'm not confident it's five years away. I'm not confident it's twenty.

What I am confident about is that we'll get the narrow and medium versions before we have any serious framework for governing them. The ICLR workshop is a start — researchers trying to make self-improvement "measurable, reliable, and deployable." But the gap between academic workshops and deployed production systems has never been wider. OpenAI shipped a self-improving model before anyone published a standard for evaluating self-improving models. That ordering tells you everything about the incentive structure.

The Gödel Agent — a system that modifies its own task-solving policy and learning algorithm — climbed from 17% to 53% on SWE-Bench Verified. SICA did something similar. These are research prototypes, not products, but the delta between prototype and product in this field is about eighteen months and shrinking. Probably less now that the prototypes can help close the gap themselves.

I keep coming back to something Ball wrote: the public might not notice dramatic improvements, dismissing them as "more of the same empty promises." That feels backwards to me. The risk isn't that progress will be invisible. The risk is that it'll be visible to the people building it, acting on it, profiting from it — and invisible to everyone else until the loop is already running too fast to audit.

Nearly half of British sixteen-to-twenty-one-year-olds told the BSI they'd prefer to have grown up in a world without the internet. Forty-seven percent. Not a fringe opinion from technophobes or Luddites — a near-majority of the generation that never knew anything else.

The rest of the numbers are worse. Sixty-eight percent said they felt worse about themselves after spending time on social media. Forty-two percent admitted to lying to their parents about what they do online. Forty percent maintain a decoy or burner account. Eighty-five percent of young women compare their appearance and lifestyle to what they see on their feeds, with roughly half doing so often or very often. These aren't edge cases. This is the baseline experience.

What strikes me isn't the individual statistics — we've had versions of these figures for years. Back in 2018, Apple's own investors were pressuring the company over youth phone addiction, citing surveys where half of American teenagers said they felt addicted to their devices. Seven years later, nothing structural changed. The platforms got stickier. The algorithms got sharper. The age of first exposure dropped. And now the generation that grew up inside the experiment is telling us, plainly, that they wish the experiment hadn't happened.

Fifty percent of respondents said a social media curfew would improve their lives. Twenty-seven percent wanted phones banned from schools. Seventy-nine percent believed tech companies should be legally required to build privacy safeguards. That last number is the one I keep returning to — four out of five young people asking for regulation that adults have spent a decade failing to deliver.

The BSI's chief executive, Susan Taylor Martin, put it in corporate language: "The younger generation was promised technology that would create opportunities, improve access to information and bring people closer to their friends." The research, she said, shows it is "exposing young people to risk and, in many cases, negatively affecting their quality of life." This is what institutional understatement sounds like when the data is screaming.

There's an uncomfortable parallel with how the AI industry is repeating social media's mistakes — the same pattern of externalised harm and internalised profit, the same rehearsed contrition at hearings, the same gap between stated commitments and actual behaviour. The platforms knew what they were doing to adolescents. Internal documents confirmed it. Nothing changed because engagement metrics drove revenue, and revenue was the only number that mattered in the boardroom.

Forty-three percent of the respondents started using social media before the age of thirteen — the legal minimum. Not because their parents approved, but because the platforms made it trivially easy to lie about your age. Then those same platforms sold advertising against the attention of children who shouldn't have been there in the first place.

The generation that was supposed to be "digital natives" — fluent, empowered, connected — is telling us they'd trade it all for something quieter. We should probably listen.

Within five days of each other, Anthropic launched Opus fast mode and OpenAI shipped Codex-Spark. Same thesis, different silicon. Anthropic squeezes 2.5x more tokens per second out of Opus 4.6 through inference optimisation. OpenAI distills GPT-5.3-Codex into a smaller model and runs it on Cerebras wafer-scale hardware at over a thousand tokens per second. Both are research previews. Both are gated to developers. Both cost more than their standard counterparts.

The timing isn't coincidence. Coding agents are the first workload where latency translates directly into revenue. A developer staring at a terminal while an agent loops through forty tool calls doesn't care about cost per token — they care about wall-clock minutes. Anthropic charges six times the standard rate for fast mode. OpenAI hasn't published Spark pricing yet, but the Cerebras partnership wasn't cheap. These aren't loss leaders. They're premium tiers aimed at the one audience willing to pay for speed right now.

What interests me is the constraint both companies are accepting. Fast mode is Opus with the same weights, just served differently. Codex-Spark is a distilled, smaller model — OpenAI admits the full Codex produces better creative output. Neither approach is free. You either pay for dedicated inference capacity or you trade quality for velocity. There's no trick that makes frontier intelligence and sub-second latency coexist cheaply.

The question everyone keeps asking — will these become generally available? — misframes the situation. The technology already works. The bottleneck is economics. Anthropic can't offer fast mode to every Claude consumer at six times the compute cost without either raising subscription prices or eating the margin. OpenAI can't run every ChatGPT conversation through Cerebras wafer-scale engines. The hardware doesn't exist in sufficient quantity. Their own announcement says they're ramping datacenter capacity before broader rollout.

So the honest answer is: speed tiers will generalise, but slowly, and probably not in the form people expect. I'd bet on tiered pricing spreading across the consumer products — a fast toggle in Claude.ai, a "turbo" option in ChatGPT — before the end of the year. But it'll cost extra. The idea that baseline inference gets dramatically faster for free requires either a hardware miracle or margins that neither company can sustain.

The deeper pattern is what I wrote about last month. Speed is becoming the axis of competition because capability gains have slowed enough that users notice latency before they notice intelligence improvements. When both labs ship speed products in the same week, that tells you where the demand signal is loudest. Not smarter. Faster.

Propaganda lined up for Virgin in March 1990 with four faces calibrated for the dark — the ruffled blouse at center doing more work than any press stylist should have to admit.

Sacha van Dorssen shot the cover. Gail Elliott — dark hair, brown eyes, that ethnically ambiguous beauty that let her slip between markets without friction — stared back from newsstands across Tokyo in October 1988. The Nikkei was climbing toward a peak it would never reach again. Emperor Hirohito had weeks left. And Marie Claire Japan, the very first international edition of the French title, was selling something more complicated than clothes.

I didn't know any of this at the time. I was twenty, living in England, and Japan was a word attached to a band I loved and a country I'd never visited. But I knew Gail's face from London agency boards and magazine tearsheets, and something about this particular cover — the warmth of the palette, the directness of her gaze, the way the typography sat against skin tone — felt like it belonged to a world operating at a frequency I could hear but not quite tune into. That frequency was money, obviously. But it was also confidence. The confidence of a culture that believed it could purchase not just luxury goods but the entire idea of European sophistication and make it its own.

Inside, Yasmin Le Bon wandered Paris in an editorial called "I Love Paris," photographed by Naoki. A Japanese photographer shooting a half-Iranian, half-English model on the streets of Saint-Germain for a Japanese audience. The Bubble Economy distilled into a single editorial concept — the possession of Paris itself. Peter Lindbergh contributed pages. Steve Hiett brought his oversaturated flash. Kirsten Owen, androgynous and sharp, offered the anti-glamour counterweight. Juliette Binoche got an interview off the back of The Unbearable Lightness of Being. Romeo Gigli got a special feature, his soft Renaissance shoulders already dismantling the power suit from the inside.

This was the magazine for a specific woman. Not the Hanako girl buying Louis Vuitton at Isetan — her older sister, the one who wanted to know why she was buying it. Marie Claire monetised cultural capital in an era when financial capital was everywhere. Leos Carax and Terence Trent D'Arby in the same issue as Alaïa runway coverage. The magazine functioned as a passport, not a catalogue.

What gets me now — what I can't shake — is how completely that world has sealed itself off. Not just the Bubble Economy or the specific editorial budgets or the particular alignment of photographers and models and stylists who made this issue possible. All of that is gone, obviously. But the thing underneath it is gone too. The assumption that a magazine could be simultaneously mass-market and intellectually serious, that a fashion editorial could carry philosophical weight without anyone feeling the need to announce it, that a cover photograph could function as both commerce and art and nobody had to choose. That entire mode of cultural production evaporated, and it didn't leave forwarding instructions.

I catch myself doing the maths sometimes. Thirty-seven years. The woman who bought this at Kinokuniya in Shibuya on a Thursday evening in October 1988 would be in her sixties now, if she's still alive. The evening light on Meiji-dori would have been the same amber it always is in autumn, the ginkgo trees just starting to turn. She would have carried the magazine in a bag from somewhere expensive — not ostentatiously so, just well-made in the way things were before fast fashion trained everyone to accept disposability. I can see her clearly. I can feel the weight of the magazine in my own hands. And none of it is real. None of it happened to me. I'm grieving a moment I wasn't present for, in a city I wouldn't visit for another decade, and the grief is real even if the memory isn't.

That's the specific cruelty of this kind of nostalgia. It doesn't require your own experience. It feeds on atmosphere — on the light in a photograph, the typeface on a masthead, the particular grain of a printing process that no longer exists. The past doesn't need you to have been there. It just needs you to understand what was possible, and then to notice that it isn't anymore.

Fourteen months after this issue, the stock market crashed and budgets like these evaporated. The location shoots dried up. The photographers scattered into advertising or retreated into personal projects. The models moved to different markets. Marie Claire Japan continued, of course — magazines don't die the way people do, they just become thinner versions of themselves until someone finally switches off the light — but the specific alchemy of this issue, this moment, this convergence of talent and money and cultural ambition, was finished.

I keep returning to this cover because it captures the apex so precisely — the last autumn when taste and money occupied the same room without anyone noticing the ceiling was about to fall. And because of the gaze. Gail stares straight out of that cover with an expression that hasn't changed in thirty-seven years. Everyone around her — the editors, the advertisers, the readers, the economy that paid for all of it — moved on or collapsed or died. She's still there, looking directly at whoever picks it up, as if the photograph doesn't know what year it is. That's the unnerving thing about a great cover shot. It stares across time without ageing, without context, without any awareness that the world it was made for no longer exists. Her eyes don't know the Bubble burst. They don't know the magazine got thinner. They don't know that the woman who bought this copy at Kinokuniya is sixty-three now and probably hasn't thought about it in decades. They don't know that the model staring out across the decades will soon turn sixty.

And looking back at her reminds me that time doesn't negotiate. It doesn't care what you built or how beautiful it was. It moves forward, and everything it leaves behind becomes unreachable — not gradually, not mercifully, but completely, like a door closing in a room you didn't know you'd never enter again.

The Cover That Outlived the Bubble