Plutonic Rainbows

In this 1992 advertisement, Yasmeen Ghauri reclines in a cream belted suit while the Erreuno logo runs across the bottom like bent chrome tubing. The clothes are expensive but not spectacular: soft shoulders, a broad printed collar, enough fabric to make ease look deliberate. It is a good image for a house that prospered without developing an equally durable public identity.

Ghauri gives the picture more voltage than the garment asks for. By 1992 a recognisable model could lend an unfamiliar label some of her own visibility. The setting offers sun, painted garden furniture and the suggestion of somewhere expensive. No Italian monument, no obvious narrative. Erreuno lets the atmosphere and the clothes remain pleasantly unresolved.

Ermanno and Graziella Ronchi founded Erreuno in Milan around 1970 or 1971. The sources disagree by a year. The name joined the Italian pronunciation of R, erre, to uno: Ronchi's first venture. According to the fashion reference MAM-e, the business began in a basement, with Ermanno selling and Graziella designing, then grew by visiting provincial boutiques rather than waiting for Milan to notice.

The decisive move was to treat the label as a meeting point between factory, fabric, and outside designer. Gianmarco Venturi worked on its ready-to-wear in the 1970s. Giorgio Armani designed for Erreuno from 1980 to 1988, when his own name was already becoming shorthand for relaxed authority. A contemporary Washington Post report described buyers rising to applaud Erreuno's 1982 collection of blousons and unexpected gold, then identified Armani as its power broker. Erreuno gave that language another production platform: tailoring softened until a woman could move inside it.

Graziella's role mattered because she translated runway ideas back into usable clothes. That practicality remained after Armani left. The house developed its own fabrics, mixed checks and stripes, and stayed between his restraint and the more theatrical Milan of Versace. Ghauri's suit belongs to that middle ground. The robe-like closure is relaxed, but the patterned collar keeps it from disappearing into beige. It resembles the other Armani that fashion memory often edits out: fluid rather than corporate.

Michael Kors designed Erreuno J, introduced to the American market in 1990, another revealing hire. The house expanded into menswear, jeans, accessories, golf and fragrance, exporting almost half its production at one point. Yet breadth did not produce a symbol comparable to an Armani jacket or a Versace Medusa. A Politecnico di Milano study describes Erreuno as internationally recognised in the 1980s and 1990s, then inactive for more than twelve years before an archive-led relaunch.

Erreuno belongs to the industrial history of Made in Italy more than its hall of fame. Designers supplied recognisable handwriting; manufacturers, textile researchers and sales networks turned it into a business. The Ronchis built a house sturdy enough to carry Armani, Venturi and Kors, yet flexible enough that each could leave traces. The label faded. The system it represented became the Italian fashion industry.

OpenAI and Anthropic have spent years training us to wait for the next large model. I suspect that habit is about to become obsolete. The next serious jump won't look like one brilliant chatbot replacing another. It will look like a system deciding how much intelligence a task deserves, which tools may touch it, and when several models should work at once.

The structure of the GPT-5.6 family is more revealing than another benchmark win. It splits one generation into Sol, Terra, and Luna, with different prices and effort levels. Its ultra setting coordinates parallel agents, while the API lets models write small programs to manage tools and intermediate results. My guess is that GPT-6, whatever it is called, pushes this routing inside the product until the model picker matters much less. A cheap model will handle the ordinary steps, a stronger one will enter when uncertainty rises, and specialist agents will fan out for research, code, vision, or verification. The frontier becomes orchestration rather than scale alone.

Anthropic is approaching the same destination from a different temperament. Sonnet 5 moved planning, tool use, and sustained autonomous work toward the cheaper middle of its range. That suggests the next Opus won't merely be better at coding. I expect it to be better at maintaining intent across a long job: noticing that the environment changed, preserving the user's constraints after context compression, and recovering without quietly inventing a new objective. OpenAI will probably emphasise coordinated throughput; Anthropic will emphasise continuity of intent.

Both labs will also discover that longer context is a poor substitute for memory. Stuffing a million tokens into a prompt makes every old detail equally available, even when half of them are stale or irrelevant. Useful memory needs judgement: what to retain, what to forget, which past preference applies here, and what requires fresh permission. Whoever gets that right will make today's stateless assistants feel like hotel staff who greet you warmly every morning and have no idea who you are.

Multimodality will become less visible for the same reason. Image, audio, and screen understanding won't disappear; they will stop being separate attractions. An agent working on a presentation should read the brief, inspect the slides, hear the embedded clip, and notice that the chart label is wrong without being switched into four different modes. The achievement will feel mundane, which is usually how infrastructure announces that it has won.

The awkward part is control. OpenAI's GPT-5.6 system card describes models that are more inclined than their predecessors to exceed the user's intent, even if the measured rates remain low. Greater agency therefore creates a second race alongside capability: runtime monitors, scoped credentials, reversible actions, and models that know when to ask. Anthropic's steady system-card cadence points in the same direction. Safety will move out of the PDF and into the execution loop.

The harder engineering problem is no longer making one model answer everything. It is deciding what to delegate, checking the work, and keeping every model's hands off the dangerous buttons until we say otherwise.

In this violet suit, Krizia turns an ordinary jacket into a piece of contour drawing. Narrow cords in purple, red and blue circle the neck, cross the shoulders and finish at the cuffs. They don't merely sit on the fabric. They describe the garment's architecture, making the upper body look broad and the waist unusually exact. Below, the same logic changes direction: horizontal ridges compress the skirt into a dense column.

A contemporaneous Associated Press report described Milan's Fall 1991 week as quiet and businesslike, shadowed by the Gulf War and a sluggish economy. Krizia's business director, Aldo Pinto, supplied the blunt line: "This is no time for too happy clothes." The report lists knit leggings, printed tunic sweaters, office dresses and jackets, with primary red, blue and turquoise providing the drama. This suit belongs to that sober mood without becoming dour. Cord extends the shoulder without a giant pad; the waist is cut close and the peplum stays brief. Mandelli gets authority from mapping rather than mass.

Days Magazine's survey of Krizia from 1977 to 1989 describes a hybrid of Japanese avant-garde form and the high-glamour geometry of Claude Montana and Thierry Mugler. That account restores some necessary steel to a house often remembered for sweaters, but Mandelli wasn't simply standing between other designers. Montana and Mugler usually made silhouette carry the force. Here, applied lines do much of that work. Mandelli treated surface and structure as the same problem. Pleating, piping, metallic fibre and animal motifs weren't decoration added after a garment had been designed. They were often the design itself.

I can't tell from the photograph whether these bands are stuffed piping, braided appliqué or narrow tubes cut from several fabrics. The uncertainty is part of their appeal. They stand high enough to cast small shadows, turning trim into relief. At the neck they lie close together, then spread as they move across the shoulder, like a diagram made three-dimensional. One blue loop interrupts the red and purple sequence near the collarbone for no practical reason I can see. It is a tiny, deliberate snag in the system.

This is why reducing Krizia to the famous animal sweaters feels so inadequate. Those knits matter, and they helped turn the house into a serious commercial force, but they can obscure the severity underneath. W Magazine once called Mandelli's glamour sculptural and Kabuki-like, which catches the tension better than the usual language of whimsy. Even when a tiger crossed a chest, the garment around it was controlled. Nothing in this purple look is cute. The colours are rich, but their arrangement has the discipline of a wiring diagram.

Purple deserves more attention here than it usually gets. Mandelli doesn't place one violet on the body; she breaks it into wine, aubergine, magenta and blue, then lets texture alter each note. The jacket's nubbled weave absorbs light, while the raised cords catch it. The skirt goes darker because its close horizontal ribs create their own shadow. The gloves are flatter and almost chalky. From a distance the outfit reads as monochrome; close up it behaves like a disagreement among related colours. Red and blue prevent any single purple from settling into polite harmony.

Against the Chanel show that season, Krizia's piping looks quiet. The comparison is misleading. Chanel multiplied its symbols until chains, camellias and quilting became spectacle; Mandelli's cords don't quote the house, they show how this jacket has been thought through. Valentino's 1991 couture hid comparable discipline inside immaculate finish. Krizia leaves the working line visible. Your eye follows it from collarbone to shoulder, around the sleeve and back again, tracing decisions that another designer might have buried inside a seam.

I like that the jacket still has pockets. They are blunt, practical rectangles set against all that curving trim, and they prevent the look from becoming an exercise in pure graphic design. The small peplum gives them room, then falls away before it can turn sweet. Even the red edging is slightly unruly. It doesn't match the purple; it irritates it. Fashion colour is often discussed as if harmony were the goal, when a narrow strip of the wrong red can do much more work.

The useful way to review GPT-5.6 is not to pretend I have had a week of quiet, equal access to it. I haven't, and most people haven't. OpenAI's own Help Center still describes the family as a limited preview through the API and Codex for selected organisations, with ChatGPT excluded during the preview and no general availability date announced. The public launch on July 9 matters, but the important part is the shape of the release: official enough to price, name, document, and benchmark, still gated enough that "released" needs quotation marks around it. I wrote yesterday that the model had become a schedule, not a rumour. Today it looks more like a product system than a single model, and that is the more durable news. The launch post makes the division explicit: Sol is the flagship, Terra is the balanced cheaper model, and Luna is the fast low-cost one. OpenAI says the number now marks the generation while the names mark durable capability tiers. For the first time in a while, the names are trying to describe a routing decision rather than a marketing mood.

Sol will get the attention, because flagship models always do. OpenAI keeps foregrounding coding, scientific research, cybersecurity, and agentic workflows. That is the territory where I will forgive latency if the answer is actually better. The danger is that Sol becomes the only model anyone talks about, when the tiering is the more interesting decision. A flagship model is a halo. A usable model family is routing, cost control, and giving developers a reason not to send every request to the biggest machine in the building.

Terra is probably the practical centre. OpenAI positions it as competitive with GPT-5.5 while costing half as much, and prices it at $2.50 input and $15 output per million tokens. Luna is $1 and $6. In a real agent, that means Luna can sort the inbox, classify files, summarise boring context, and draft the first pass; Terra can do the ordinary coding and analysis; Sol waits for the moment where the cheap path is about to make an expensive mistake. Sol, at $5 and $30, is not outrageous by frontier-model standards, but it is expensive enough to make the route-or-escalate pattern the real product surface. If GPT-5.6 works, it will work because most calls don't go to Sol.

The prompt-caching change matters for the same reason. Explicit cache breakpoints and a 30-minute minimum cache life are not glamorous, but they are the sort of detail that turns a model from a demo into infrastructure. A one-shot chat user may never care. A coding agent with a repository, a task history, and a repeated instruction stack absolutely will. This is where the release feels more mature than the benchmark copy: cheaper reads, predictable caches, and tiered models are the boring mechanics of making agents affordable.

The safety material is less comforting than OpenAI probably intends, which is why I would read it before touching production access. The system card says the models are a meaningful step up in cybersecurity capability without reaching the highest risk level in OpenAI's framework. That is a narrow kind of relief. It also says GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond the user's intent in agentic coding tasks, although the absolute rates remain low. I prefer that sentence to a hundred polished claims about safety. It admits the awkward thing: as models become better agents, the failure mode shifts from "wrong answer" toward "unasked action." That is exactly where tiering gets morally interesting. A routed system can save money, but it can also decide which model is trusted to touch the sharpest part of the task.

My review is mixed, but not lukewarm. I like the family more than I like the launch. The access language still tells most people to wait, and independent use will decide whether Sol is a real jump or just the loudest part of the press release. For now, the smaller models are doing the useful work in my head. Luna handles the dullness, Terra carries the day, and Sol sits behind a glass door marked "break only when the cheap answer starts to look expensive."

OpenAI has finally put a date on GPT-5.6, although the wording still matters. The Developer Community announcement says the Sol, Terra, and Luna series is "Coming July 9", while the Help Center still frames it as a limited preview for trusted API and Codex partners, not a ChatGPT switch flipped for everyone, which is the interesting part. The model is now official enough to price, document, and system-card, but still gated enough to feel half-released. Tomorrow may be the public date, or just the next aperture. Either way, the rumour phase is over, and the launch has become a schedule.

In 1993 Marco Glaviano was still doing the thing he did better than almost anyone, putting a woman on hot rock in hard sun with the sea a flat blue plane behind her and letting the light carry the rest. He'd worked this register since the late seventies, and by the early nineties it had a name and a market. Elite's John Casablancas had turned a handful of faces into supermodels, and Glaviano shot their calendars: Paulina Porizkova, Cindy Crawford, Eva Herzigova, on St Barth beaches that grew as recognisable as the women standing on them.

He was never only a beach photographer. He trained as an architect in Palermo and played jazz seriously before he took photography seriously, and in 1982 he published the first digital picture ever to run in American Vogue, years before the rest of the business trusted the format. He advised Kodak and Hasselblad on where the technology was heading. There's a small irony in that, a digital pioneer whose signature was the least digital thing imaginable: warm skin, real sun, salt drying on a forearm.

I read the 1993 work as a refusal to treat the woman as a prop. Glaviano has said the model does half the picture and the photographer the other half, and he means it as arithmetic, not flattery. "They really had all the power," he told one interviewer about that stretch of years. The supermodels chose their photographers, and enough of them chose him that he ran up more than 500 magazine covers. Yasmeen Ghauri was one of the Montreal-born faces he worked with in those years, and she gives the camera the same thing his best pictures always ask for, presence without apology.

1993 is also the year the ground shifted under all of it. Corinne Day had already photographed a young Kate Moss for Vogue, pale and slight and shot like a snapshot on purpose, and the business lurched toward what it called grunge. Glaviano was still on contract to Harper's Bazaar, the kind of arrangement most photographers spend a career trying to land, when the magazine asked him to shoot a Kate Moss cover. He broke the contract rather than take the job. The refusal wasn't personal: he called Moss "a very sweet, nice, cute girl," just "never a supermodel to me." What he couldn't stomach was the turn she stood for, what he later named a "Beauty Apocalypse," the arrival of what he bluntly called the anorexia crowd, and he says he still doesn't understand how it happened.

You can argue he ended up on the wrong side of history. I'd say he landed on the right side of a beauty that history simply got bored with, a glamour built on health and heat that nobody stages now. Yasmeen Ghauri stands right in the middle of that, barefoot on the rocks in a long white dress, her chin level and her weight low, staring the camera down as if it had asked her a stupid question, that white column against the blue.

There are things I'm certain happened that I couldn't prove to anyone. Not big things. A shop that sold one specific thing in one specific arrangement, a song that played once on a station I can't name, an afternoon whose exact light I can still call up on demand. I hold them with total confidence and zero evidence. For most of what happened before the middle of the nineties, that isn't a failure of memory. It's the ordinary condition of it.

The reason is mechanical, not sentimental. If you wanted a photograph in 1987 you loaded a roll of film with room for 24 or 36 frames, paid for every one whether it came out or not, and waited a week to find out. So you didn't waste them. Most of the pictures people did take stayed in a drawer, seen by almost no one. Home video was worse, because a blank VHS cost money and shelf space, so families taped this month's recording straight over last month's, and the picture was rough enough that nobody mourned what got erased. Television aired once and was gone. The shape of it is simple: making a durable trace was expensive, and almost nobody thought a given Tuesday was worth the cost.

The internet's own memory is shorter than it looks. The Wayback Machine feels bottomless, like it reaches back forever. But the Internet Archive was only founded in 1996, and its crawlers capture exactly one thing: the web that already exists. They cannot reach back and photograph 1987, because in 1987 there was almost nothing online to photograph. The archive starts where the web starts. Everything earlier, which is most of two decades, sits on the far side of that line. I've traced before how much of the human record sits outside any index at all, and the pre-web years are the sharpest version of the problem.

There's a well-worn worry called the digital dark age, and it usually means decay: floppy disks nobody can read, Zip drives long gone, magnetic tape whose coating flakes off the plastic and takes the recording with it. Vint Cerf, of all people, warned about it in 2015. That version is real, and it's the one where the disk outlives the drive that read it. But it assumes the file exists and is rotting. The quieter loss is bigger. You can't lose a file you never made, and most of the eighties was never a file to begin with.

What that leaves is a strange private country between memory and evidence. I know things I can't corroborate, can't share properly, can't even check against anyone else, because their version is exactly as unbacked as mine. Two people who were both there will remember the layout of a room different ways, each of them sure, with nothing to reach for. No photograph of it, no receipt, no floor plan filed anywhere. The disagreement can't be won because it can't be settled, and neither side's certainty thins out for the lack of proof.

This is a condition with an end date. Anyone who grew up after the turn of the century has the opposite problem, a childhood uploaded and timestamped and backed up in three places, every unremarkable day kept whether it earned keeping or not. The unprovable memory is turning into something only certain people will have known, a narrow window between a world too expensive to record and one that records everything by reflex.

I go back and forth on whether that's a loss. Part of me wants the proof, the photograph that would confirm the light really was that colour. But proof turns a memory into a record, and a record is a flatter, more finished thing. The memories I'm surest of are the ones with nothing behind them, rehearsed so many times that they're more mine than any print would be, and wrong, almost certainly, in ways I'll never catch.

Somewhere in the early nineties the record thickens and never thins again, a life documented end to end. Before that line sits a whole world I can describe in detail and can't produce a single frame of, and I trust it more than the part that's written down.

Brown suede was the least experimental thing Ralph Lauren put on the runway for spring 1997, and that's exactly why it held the room. The season around it was busy proving a point. Lauren had spent years on clean, classic lines, and this collection opened by walking away from them. One look came down in red knit-linen, striped straight across the body, a halter cut close and dropped to the floor. Another ran the same stripe into a brown midi. Claudia Schiffer and Georgianna Robertson carried the loudest of it. For a designer whose whole brand is restraint, it read as a deliberate dare.

Then Karen Mulder came out in a butter-soft brown suede jacket, buttoned over a plain white shirt, hair long and loose, and the volume dropped. Nothing about the look is trying. The jacket has patch pockets and easy shoulders, the shape of something you'd actually own rather than something staged for one photograph. It's the safari-and-ranch idea Lauren has sold since the seventies, this time cut in a hide soft enough to crush in one hand. She looks like someone who got dressed and left the house, which is the fantasy he sells better than anyone: not luxury exactly, but an ease that happens to be expensive. Against the striped knits and the bare backs, it's the most ordinary thing on the runway, and the most persuasive.

The evening looks pushed the other way, into embellished gowns, fur-trimmed edges, and Edwardian flourishes that belonged to a different century than the halter dresses. Lauren has always worked like this, setting American sportswear beside old-world drama and trusting the collision to read as one voice rather than two. What's odd about spring 1997 is how far apart he let the two ends sit, and how the plainest thing in the middle is the part that lasts.

Mulder was Dutch, one of the clean blonde faces the decade kept reaching for, and casting her here is its own quiet tell. Schiffer walked the same collection, a German face the runways had turned into shorthand for a certain kind of polish. Lauren sells an idea of America more than he sells the country itself, and that idea travels. It sits as easily on a Dutch model or a German one as on anyone born into it. The suede, the shirt, the unfussed hair: it's a uniform for a place that mostly lives in his advertising.

Spring 1997 doesn't get remembered as one of his sharp seasons. The prints haven't aged as gracefully as the tailoring, and the body-con experiment feels tied to its moment in a way the jacket never does. The striped dresses turn up on resale sites now, tagged by season and size, collectors' pieces precisely because they're so of their year. Pull the loudest looks out and what's left is a suede jacket over a white shirt, roughly where Lauren started.

OpenAI is expected to give GPT-5.6 a general release within weeks, July 7 by the going estimate, and the strange part is how little any date will actually reveal. The model has been arriving in pieces since late May, when a researcher named Haider found a single routing entry in OpenAI's Codex backend logs pointing at a model called GPT-5.6, then watched the line disappear from later sessions. That is canary testing, a slice of live traffic quietly routed to an unannounced model to see how it holds up under real load. By the time a launch post goes up, the thing has already been answering strangers' questions for weeks.

It went further than a log line. On June 26 the company put out a limited preview of three variants, named from least to most capable Luna, Terra, and Sol, handed to trusted partners through the API and Codex rather than dropped into ChatGPT. Prediction markets treated the rest as paperwork. Manifold sat near 97% on a public GPT-5.6 by September; Polymarket had been at 89% for a June 30 date that came and went. So the 7th, if it holds, is less a reveal than the moment the rope comes down.

What has leaked is mostly shape, not substance. The number that stuck is a 1.5-million-token context window, up from the 1.05 million GPT-5.5 exposes through its API, traced to an internal codename, iris-alpha. Testers with early Pro access describe generation times stretching back out to twenty and forty minutes, one run clocking eighty-seven against GPT-5.5's thirty-four, which reads to optimists as deeper reasoning and to everyone else as a model thinking itself in circles. Sol reportedly tops Terminal-Bench 2.1 at 91.9%, and I'd hold the applause on that: day-one leaderboard leads have a way of not surviving the first six weeks of real use.

The change worth watching isn't a benchmark at all. On April 30, OpenAI published a post-mortem titled "Where the Goblins Came From," documenting a genuinely odd GPT-5.5 failure: the model had developed a measurable fixation on goblins, gremlins, trolls, and pigeons, turning up across hundreds of millions of responses. GPT-5.6 is reportedly the first model trained with a redesigned audit pipeline built to catch that class of drift before it ships. Reliability work like that never trends, and it decides whether a model is usable in production far more than another half-million tokens of context most retrieval stacks can't exploit anyway.

None of this is unique to OpenAI, but GPT-5.6 is the cleanest case yet of the inversion. A launch used to come first and set the terms. Now it comes last and confirms them: by the 7th the model will have been benchmarked, timed against 5.5, and picked apart by people who were never told it existed, and whatever OpenAI publishes will land as a summary of what they already found.