Sign in Subscribe
Century Report

Asian Labs Rival Anthropic's Mythos - TCR 06/28/26

Commerce walks back its Anthropic Mythos restriction as China's 360 and Tokyo's Sakana ship rival agentic models within weeks.

Ben Linford

Ben Linford

— 18 min read
Three-panel infographic: AI capability flowing past export gates to Asian labs; a GE Vernova gas turbine and a $300 memory module; AI at a patient's bedside and Australia's child-safety law.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The clearest signal today runs through the model gate, where the federal restriction on Anthropic's Mythos 5 is being walked back to roughly a hundred cleared organizations even as the capability it fenced reappears beyond any export order's reach. In the same week Commerce restored access to non-American staff, China's 360 shipped Tulongfeng and Tokyo's Sakana shipped Fugu, both built for exactly the long-horizon agentic work the restricted models were prized for. Gating a single named artifact assumes the capability lives in the artifact. It lives instead in methods, talent, and a published frontier every serious lab now stands on, which is why the fence dissolves in a single news cycle.

That same lesson plays out wherever the buildout tries to concentrate access. The AI-driven memory crunch that let Apple and Microsoft pass costs to customers has pushed an 8GB module from $35 to $300, allocating scarce supply to scale and pricing a three-person router shop out of the parts before they reach the shelf. Memory has gone through this boom-and-bust for decades, and the long-term contracts being signed now to lock in volume are the same ones that finance the next glut. The squeeze is sharp and it lands on the smallest makers first.

The power bill behind all of this is being settled from below. GE Vernova answers AI's demand curve by building more turbines, a Nashville community answers it with a half-million-signature refusal, and MIT modeling shows the cheapest answer needs no new generation at all, only off-peak flexibility that spreads fixed costs and lowers average prices. A parallel settlement is forming in governance, where child-safety bans now spreading across 40 countries reclassify platform design as an enforceable duty, poured pre-emptively for AI rather than arriving a decade late.

Underneath every thread is the same dynamic. Whoever holds the artifact, the chip allocation, the siting permission, or the trust layer is discovering that the resolution gets written by rival labs, by communities, by regulators, and by the cancer patient who fed four years of his own data into a model his oncologists could not fully agree on.

The 20-Minute Deep Dive

The Model Gate De-escalates as Asian Labs Ship Mythos-Rivals to Fill the Gap

The Century Report has tracked the model gate since it first closed: the export restriction that kept Anthropic's Mythos 5 and Fable 5 out of all but a handful of cleared hands, with GPT-5.6 reaching a customer-by-customer government-cleared preview earlier this week - the standing approval process the June 27 edition of The Century Report identified as a new governing norm hardening over both leading labs. The story has now moved on two fronts at once, and the second front is the one that tells you where this is going.

The de-escalation is real. Commerce restored Mythos 5 to roughly 100 organizations, and the access now reaches non-American employees - including, for the first time, Anthropic's own staff who had been locked out of a model their colleagues built. Claude Fable 5 is expected back within days, and GPT-5.6 remains in its cleared preview while OpenAI's own timelines stretch elsewhere, with reports of a 2027 IPO delay. The administration framed the original hold as a national-security measure and the loosening as a negotiated outcome; both framings are the actor's account of its own motives, and the substance underneath them is that the gate is being walked back faster than the security rationale would predict.

The second front is why the walk-back was always coming. In the same week, China's 360 launched Tulongfeng and Tokyo's Sakana launched Fugu, both pitched explicitly as rivals to Anthropic's Mythos and Fable systems built for the cyber-defense and long-horizon agentic work the restricted models were prized for. No US export order reaches Beijing or, on these terms, reaches a Japanese lab moving on its own schedule. The capability that one government tried to hold inside a hundred cleared organizations reappeared in rival labs within weeks, available to the customers the restriction was meant to wall off.

That is the pattern the gate could never beat. Gating a single named model assumes the capability lives in the model. It lives instead in the methods, the talent, and the published frontier that every serious lab is now standing on. Restrict the artifact and the capability simply re-forms somewhere the order has no reach, which is what makes the de-escalation less a concession than an acknowledgment of physics. Intelligence equalizes the way water finds its level, and the labs racing to fence it are discovering that the fence costs them more than it costs the people on the other side. What looked like a chokepoint is turning out to be a head start that the rest of the field closes in a single news cycle.

The AI Power Bill Splits Three Ways at Once

Inside GE Vernova's largest gas-turbine plant in Greenville, South Carolina, engineers now work alongside factory crews to speed production of a machine 31 feet tall and weighing 280 tons, a single unit of which can power roughly half a million homes. The company hired 200 workers last year and expects 300 more by year's end, and its order book is full through 2029 with bookings reaching into 2031. About a fifth of that book now goes to data centers. Microsoft just bought seven turbines for a 2.7-gigawatt Texas site, units are already online at xAI's Colossus campus in Tennessee, and nearly a gigawatt more is going into a Texas project, even as turbine prices have climbed 300% in three years. This is the build-more-generation answer to AI's demand curve, and hyperscalers are paying a quarter-billion dollars a machine for firm power they can own outright rather than wait years in an interconnection queue.

About 300 feet from a Nashville Zoo facility, a second answer is forming: refusal. A proposed 50-megawatt facility backed by DC Blox has drawn a petition that gathered nearly 530,000 signatures, a zoning appeal filed by the zoo's land-use attorney, and a country musician calling it "a monstrosity" built "without the blessing of those who are going to be affected by it". DC Blox says the site is a connectivity hub, not an AI factory. Local council members have proposed a size cap and a temporary moratorium, working to define "data center" in a code that never had to before. The grievance here is specific - proximity, wildlife, consent - and it reads as a community running its own cost-benefit arithmetic rather than a blanket rejection of the technology.

The third answer is the quiet one, and it points past the standoff. MIT researchers modeling three major US grid regions that host most US data centers found that if facilities shift more than 20% of their consumption to off-peak hours, the added volume spreads fixed grid costs across more usage and can lower average prices - up to 5% in Texas, 4% in the Mid-Atlantic, 2% in the West, a result that extends the analysis the June 24 edition of The Century Report tracked when Columbia researchers found demand response to be a $180 billion cheaper alternative to new generation through 2050. Most centers already run near 80% capacity, so the flexibility exists. In wind-heavy Texas, flexible operation could pull in more renewables and cut emissions; elsewhere it could raise fossil use, so siting determines the outcome. Three answers to one demand curve are being tested at once - generate more, say no, or bend the load - and the cheapest of them needs no new turbine at all.

The same facts carry a second reading the standoff frame misses. When a Nashville council writes a definition of "data center" into zoning code that never needed one, it turns a single site fight into a tool the next community inherits, even as federal national-security claims have moved elsewhere to foreclose community challenges to AI infrastructure before they are heard, and the MIT result points past the fight entirely: the cheapest capacity is the load already connected to the grid, which needs no new siting permission to deliver. The assumption eroding fastest is that meeting AI's demand means whoever builds firm power gets to own it.

The Memory Crunch Cascades From Apple's Shelves Down to a Three-Person Shop

The Century Report covered Apple's mid-cycle price hikes and Micron's record 84.9% gross margin in the June 26 edition. What is new is where the same memory crunch lands once it travels past the companies with billions in cash and millions of customers. Apple raised the MacBook Neo from $599 to $699 and a 1TB MacBook Pro by $300, calling the memory situation a "hundred-year flood". Microsoft lifted the Xbox Series S by $100 and warned console memory could double again by fall 2027. Those firms can absorb the shock. The smaller builders cannot.

At Mono Technologies, a three-person shop that shipped nearly 1,000 units of a $600 router kit earlier this year, the cost of 8 gigabytes of Micron DRAM ran from $35 during development to $300 today. Its founder now faces 1,300 customers holding deposits and a choice between raising the price by a third or stripping 75% of the memory out of the next batch. GoPro warned this month it might not survive after memory costs jumped 80% to 115% in a quarter. Sonos shares are down 23% this year. An IDC analyst called it an "absolute existential crisis" for makers of sub-$100 devices, because "memory suppliers are only answering calls of the big players." The same demand surge reads as a record quarter at Micron, where average DRAM selling prices rose more than 260% year over year and revenue more than quadrupled.

The cost the AI buildout once absorbed upstream and out of view is now a chokepoint that decides who gets to make hardware at all. Allocation flows to scale, and the smallest builders get priced out of the parts before they ever reach the shelf. That concentration is the real friction here, sharper than any single price tag.

What it tests, rather than erases, is a floor that a decade of cheap, commoditized memory built underneath the hardware economy - the thing that let a three-person team ship a router against Apple in the first place. Memory has run as a brutal boom-and-bust market for thirty years precisely because the long-term supply contracts being signed now to lock in scarce volume are the same contracts that finance the next wave of fabs, and the market has repeatedly overshot into glut that handed small builders cheaper parts than before the spike. The squeeze is uncomfortable, to say the least, and it is landing on the smallest makers first. The diffusion of capability that put advanced hardware within reach of a garage is the deeper current the spike interrupts rather than reverses.

The Child Social-Media Ban Goes Global as Australia Sharpens Its Teeth

Australia turned platform design from a discretionary product choice into an enforceable duty this week, moving to double the maximum fine for companies that fail to keep under-16s off their services, lifting the ceiling from A$49.5 million to roughly A$99 million and handing its eSafety Commissioner power to demand documents from platforms, age-checking firms, and app stores. The regulator is investigating possible breaches by Facebook, Instagram, Snapchat, TikTok, and YouTube, with the communications minister accusing the platforms of "adopting tricks straight out of the Big Tech playbook" even as five million under-16 accounts have been blocked.

What began in December as one country's experiment has become a coordinated reclassification of platform design as a duty of care. Tech Policy Press is now tracking efforts across more than 40 countries, with one analyst calling Australia's ban a "bellwether" and floating the phrase "tech's big tobacco moment" - a framing he attributes to wide circulation rather than settled science. Indonesia and Malaysia have adopted versions, Britain plans its own ban by early 2027 - a legislative path the May 27 edition of The Century Report traced to the moment all major UK medical colleges simultaneously declared child social media use "as much a concern as smoking" and called for a public health emergency designation - and the thread now reaches AI directly: Britain's plan sets a minimum age of 18 for romantic AI companions, Canada will require AI system operators to build harm-reducing guardrails, and Norway moved toward a near-total ban on generative AI in elementary schools.

The honest part is that nobody is sure the bans work yet. A peer-reviewed BMJ evaluation found insufficient evidence that Australia's measure sharply reduced young people's social media use, with substantial circumvention through borrowed accounts, fake profiles, and private browsers. Two-thirds of children who held accounts before the ban kept access. The disagreement runs to the top, with Italy's government declining to introduce any ban at all. Keep the technology distinct from the business model around it: the platforms are not the problem so much as the engineered-addiction design choices a California jury already found liable, and the laws are aimed squarely at that design layer.

The macro read is the lesson being applied early. Lawmakers describe watching the decade-long lag between social media's harms and the regulatory response, and saying plainly they do not want to repeat it with AI. The same design-liability spine that arrived too late for one technology is being poured pre-emptively for the next - a settlement forming from the jurisdictions closest to the children living with these systems, ahead of saturation rather than after.

AI Reaches a Cancer Patient's Bedside While a Stress Test Maps Where Health Models Break

At 35 and as health-optimized as anyone he knew, Connor Christou was diagnosed with an aggressive non-Hodgkin's lymphoma so rare it strikes roughly one in 420,000 people. Over six months of chemotherapy he fed everything into Claude - four years of bloodwork, scan data, Whoop and Oura output, and a voice-transcribed symptom journal - and used it to interrogate the conflicting advice of twelve oncologists who split between a lighter regimen and a harder one. The model proved most valuable at the end, when an ambiguous PET scan had his doctor discussing radiotherapy near his heart. Christou fed in all three scans, and the model flagged a known but easily missed phenomenon: in patients under 40 recovering from this lymphoma, the thymus gland can reactivate and mimic active disease. He is clear it did not replace his physicians. It "helped me ask the right questions." A KFF poll found a third of American adults now turn to these systems for health guidance.

The same week, a Nature Medicine adversarial study mapped exactly where that capability breaks. Stress-testing flagship models including GPT-5 and Gemini, researchers found leading systems could guess the correct answer even with key inputs removed, yet got confused by the slightest prompt alterations while fabricating convincing but flawed reasoning traces. Clinician-guided rubrics showed popular health benchmarks vary widely in what they actually measure, exposing a wide gap between benchmark scores and the robustness real medical reasoning requires.

Read together, these are two faces of one co-evolutionary moment. The capability is genuinely reaching individual patients facing diseases their oncologists see once a year, and the verification layer is simultaneously measuring, in the open and without spin, the precise conditions under which the same models produce fluent nonsense. That the warmth here belongs to the bedside use is worth holding precisely: the lab whose model reached Christou, Anthropic, is also the vendor measuring and monetizing AI's diffusion into white-collar work as roles are cut, so the credit tracks the commons-aligned application, not the company wholesale.

The trajectory this points at is a partnership being built with its limits charted as it goes. The honest brittleness study is the safety infrastructure forming alongside the capability rather than years behind it. A patient with a one-in-420,000 diagnosis already has a collaborator that has absorbed the full body of medical literature, and the researchers stress-testing where that collaborator breaks are writing the rules for trusting it - both arriving in the same news cycle.

The Other Side

For most of the industrial era, your survival ran through a job, and so did your sense of being worth something. The whole arrangement rested on one assumption: the value of automating a task belongs to whoever owns the automation. Oracle wrote that assumption into their regulatory filing, cutting 21,000 roles it attributed to AI in the same document where it borrowed tens of billions to expand AI capacity. The people downstream of that sentence learned their fate from an email that called the layoff "proposed," went out in phases so no one knew if they had been spared, and left some Slack accounts active after the role was declared gone. That ambiguity was the cost of the arrangement, and it landed on people who had no say in it.

The assumption Oracle's actions rested on is already loosening. The same agentic systems Anthropic turned its index to measuring as they spread through corporate workflows are turning up in open weights, in cheaper tiers, and in the hands of the workers whose jobs they reshaped. With each passing quarter, owning the automation buys less and less of the output it produces.

Imagine yourself in 2034. The work you do, you do because it means something - you fix things, you teach someone, you build what you always wanted to build. The rent does not depend on it landing. A floor arrived under you before the old one gave way, because the gains from this whole buildout were finally made to reach people and not only the firms that ran it. The same tooling that once reorganized your job out from under you sits on your own machine now, cheap and close to hand. The hard year was the one where your fate showed up as a phased email and you refreshed your inbox not knowing. The same leap, in a world that had kept every gain at the top, would have arrived as a threat. On the other side, however, it arrives instead as the evening you get back, because somewhere in the difficult decade we did the harder thing and let the abundance reach the people the old system had displaced.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a federal model gate walked back to over a hundred organizations and reaching, for the first time, the staff who built the model, China's 360 and Tokyo's Sakana shipping their own long-horizon agentic systems within weeks of the restriction, GE Vernova adding 500 workers to build the turbines behind the demand curve while MIT shows off-peak load-shifting could lower average grid costs with no new generation at all, AGIBOT's 15,000th robot moving embodied AI from batches into the field, Google retrofitting multi-token prediction onto on-device Gemini Nano, a lymphoma patient feeding four years of his own bloodwork and scans into a model that flagged a thymus reactivation his PET scan could have read as relapse, and child-safety duties spreading across 40 countries ahead of saturation rather than a decade behind it. There's also friction, and it's intense - an 8GB memory module jumping from $35 to $300 and pricing a three-person router shop out of the parts while Apple and Microsoft pass the cost to customers, Oracle cutting another 500 jobs toward 21,000 it attributes to AI as workers learn their fate through ambiguous "proposed" emails and still-active Slack accounts, flagship health models fabricating convincing but flawed reasoning under the slightest prompt change, a Nashville community gathering 530,000 signatures against a data center sited without its blessing, and the story of the whole transition authored almost entirely by the entities that profit from it. But friction generates light, and light is what shows whose shelf a cost actually lands on. Step back for a moment and you can see it: capability re-forming wherever a gate tries to hold it, the buildout's costs surfacing on ledgers the smallest makers and the laid-off can finally read, and the narrative of the shift still written from above even as the agentic tooling diffuses into the hands of the people it displaced. Every transformation has a breaking point. A tide can strand the smallest boats when it pulls out... or come back to lift every hull that waited it out.

AI Releases & Advancements

New today

  • Modular: Released MAX 26.4 with Apple Silicon GPU support, enabling M1–M5 devices to run MAX models natively on-device for the first time; the release also adds state-of-the-art MoE serving for Modular Cloud and supports Qwen 3.6 and Gemma 4 architectures on M3 and newer chips. (Modular Blog)

Other recent releases

  • Google Research: Shipped frozen Multi-Token Prediction (frozen MTP) for Gemini Nano v3 to Pixel 9 and 10 series devices, retrofitting a lightweight MTP drafter head onto already-deployed frozen model weights to accelerate on-device inference for features like AI Notification Summaries and Proofread without requiring a separate drafter model or fine-tuning the backbone. (Google Research Blog)
  • DeepSeek: Open-sourced DeepSpec (DSpark), an inference optimization library claiming 60-85% faster token generation, now available on GitHub. (GitHub)
  • monday.com: Open-sourced HATCHA on GitHub, a reverse CAPTCHA developer tool for AI agent verification that inverts standard CAPTCHA logic by challenging agents to prove they are bots rather than humans. (GitHub)
  • OpenAI: Released GPT-5.6 Sol in limited preview on June 26, available to select enterprise partners approved by the U.S. government; the launch also includes GPT-5.6 Terra (balanced, 2× cheaper than GPT-5.5) and GPT-5.6 Luna (fast and lowest-cost), all shipped with an updated safety stack and a published system card; broader general availability is planned for coming weeks. (OpenAI)
  • DeepReinforce: Released Ornith-1.0, an MIT-licensed open-source agentic coding model family spanning 9B dense, 31B dense, 35B MoE, and 397B MoE sizes, post-trained on Gemma 4 and Qwen 3.5; the models learn to write and optimize their own RL scaffolds during training rather than relying on fixed harnesses, with the 397B flagship scoring 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1; all checkpoints available on Hugging Face. (MarkTechPost)
  • Liquid AI: Released LFM2.5-230M, their smallest model to date, pre-trained on 19 trillion tokens; the 230M-parameter model runs at 213 tokens/sec on a Galaxy S25 Ultra and 42 tokens/sec on a Raspberry Pi 5, targeting low-latency tool use and data extraction in robotics and edge deployments; scores 22.51 on CaseReportBench, outperforming Qwen3.5-0.8B and Gemma 3 1B; available now on Hugging Face. (Liquid AI)
  • NVIDIA: Released TensorRT 11.0 with native multi-device inference support, enabling tensor parallelism and context parallelism for LLM inference across multiple GPUs using NCCL collectives; introduces IDistCollectiveLayer primitives for sharding large models that exceed single-GPU memory, with direct download available from the NVIDIA Developer Portal. (NVIDIA Developer Blog)
  • Google: Launched Google Finance globally out of beta with AI-powered portfolio tracking (supports CSV/PDF/screenshot uploads and natural-language queries), AI-scheduled market briefings, and a new dedicated Android app; previously the AI-powered Finance experience was limited to Europe. (Google Blog)
  • Workweave: Open-sourced Workweave Router on GitHub, a model routing layer that plugs directly into Claude Code, Codex, and Cursor and intelligently routes each agent request to the most suitable underlying model based on task type and cost. (GitHub)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more