Sign in Subscribe
Century Report

Frontier AI Dips, Open AI Rises Worldwide - TCR 06/15/26

Reporting tied Anthropic's worldwide shutdown to a China-access fear as the AI frontier benchmark fell for the first time, and open AI surged in answer.

Ben Linford

Ben Linford

— 17 min read
Four-panel Century Report infographic: budget AI models beat the frontier as Anthropic retreats, prime-editing medicine advances, tech layoffs sit beside new wealth, states pass AI laws.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A national-security order pulled the two most capable public AI models offline on roughly 90 minutes' notice this week, and for the first time since the index began, the Artificial Analysis frontier benchmark moved backward. Yet in the same window OpenRouter showed that a panel of three budget models, synthesized by a judge, outscored both GPT-5.5 and Opus 4.8 on hard research tasks at half the cost. One state can remove a single named model from circulation. It cannot remove the capability, because the capability has already started living in the ensemble.

That gap defines the day. The export control reaches one company while rival labs stay untouched, open weights of comparable strength ship freely, and India and European governments accelerate sovereign plans in direct response. The assumption underneath every gating mechanism, that capability concentrates in a model you can name and therefore restrict, is the thing this week disproved.

The same loss of containment shows up in what is becoming visible. Tech layoffs hit their highest single month in two years with AI the most-cited reason, landing on the same ledger as IPOs minting fresh billionaires. Meta's prototype of smart-glasses face recognition, built on a Pentagon-and-police surveillance vendor's stack, surfaced because journalists read shipped code, not because a regulator forced it. The abundance is produced and then fenced, and the fencing no longer happens out of frame.

What pushes back is distributed by nature. Six states across both parties poured binding AI law one jurisdiction at a time, even as a federal preemption push aims to freeze exactly these state laws. A two-thirds Trump-voting Georgia county gathered signatures to force a referendum against an 800-acre data center. Accountability accumulates the way capability does, in many places at once, where no single point is critical.

And the frontier that genuinely widens human reach kept moving. A lipid-nanoparticle platform cleared the delivery bottleneck holding prime editing out of the clinic, correcting a phenylketonuria mutation to anticipated-curative levels in mice. AstraZeneca's Truqap became the first targeted therapy for a poor-prognosis prostate-cancer subgroup. Precision arriving as access is what the unfenced version of this transition looks like.

The 20-Minute Deep Dive

The Layoff Wave and the Wealth It Sits Beside

Tech layoffs reached their highest single month in two years, with nearly 40,000 cuts, and AI was the most-cited reason for the third month running. The pace for the year runs about 974 people per day, 44% faster than last year. What makes the moment combustible is the contrast: the displacement and an extraordinary concentration of wealth are now landing on the same ledger. Cerebras minted two billionaires on its IPO; SpaceX's first trading day, which the June 13 edition of The Century Report covered as the moment its controlling owner predicted AI and robotic labor would make money "stop being relevant," turned its owner into a paper trillionaire and an estimated 4,400 employees into millionaires; Anthropic and OpenAI are inching toward public markets near or above a trillion dollars each.

A live dispute runs underneath over whether AI is the cause or the cover. Marc Andreessen called it the "silver bullet excuse" for cuts that are really about pandemic-era over-hiring, and Block's Jack Dorsey, after laying off nearly half his company, eventually conceded he had over-hired. The honest read holds both at once: AI is genuinely absorbing work, and "AI" is also a convenient label for a correction managers wanted anyway. The worker shown the door cannot tell which one signed the notice.

The institutional response is becoming visible at the front of the pipeline as much as the back. China revoked or suspended 12,200 undergraduate degree programs between 2021 and 2025 while opening 10,200 new ones, reshaping more than 30% of the nation's university offerings, cutting arts, humanities, and languages in favor of embodied intelligence and other tech fields. This is a top-down redirection of where young people are allowed to point their training, the same national-capability strategy that this week made the country the suspected route for accessing the most capable public AI model. It steers students; it does not ask them.

Read forward, the contrast itself is the signal. For most of the industrial era, gains could concentrate unseen while displacement was absorbed somewhere out of frame. What is breaking is the quiet. The $170 million mansion and the 8,000-person layoff now sit two months and one news cycle apart, legible to anyone watching, against a cost-of-living squeeze 76% of Americans name as their top concern. The pressure that builds when abundance is produced and then visibly fenced is exactly what forces the redistribution question into the open - the universal-capital-account and guaranteed-income proposals two leading labs floated this month are early evidence the fence is already under strain.

A Delivery Breakthrough Moves Prime Editing Toward Routine Use

Prime editing is the most precise genome-editing method science has produced. It can write nearly any small change at a chosen spot in a living cell's DNA, substitutions, insertions, deletions, without the blunt double-strand breaks that older approaches rely on. The problem has never been the editing. It has been getting the three-part editing machinery into the body in a form that works, can be given more than once, and does not linger long enough to cause collateral damage. That delivery bottleneck has kept the most exact tool in the field largely confined to lab dishes and one-shot viral approaches.

A team reporting in Nature Nanotechnology appears to have cleared it. They built a systematic platform for optimizing lipid nanoparticles, the same synthetic fat-bubble carriers that delivered the mRNA in recent vaccines and the first in-body gene-editing corrections in patients, and tuned the cargo design specifically for prime editing's heavier, more complex payload. The result: 49% average prime editing in bulk mouse liver from a single low dose. In a mouse model of phenylketonuria, an inherited metabolic disorder, the system corrected the disease-causing PAH R408W mutation and pushed blood phenylalanine to levels the researchers describe as anticipated to be curative.

Three properties matter as much as the headline number. The nanoparticle approach produced fewer off-target edits than DNA-based delivery. It caused only a brief, transient rise in liver enzymes with no observed long-term toxicity. And because the carrier is synthetic rather than viral, it can be dosed repeatedly, each round building on the last, where viral delivery typically gets one shot before the immune system shuts the door.

This is demonstrated capability in mice, not a therapy a patient can request. What it moves is the date. Phenylketonuria, alpha-1 antitrypsin deficiency, sickle cell disease, several retinal disorders, these are the conditions prime editing has already corrected in animals or early patients, and the redosable non-viral route widens which of them can reach the clinic on a practical timeline. The work ahead is regulatory, manufacturing, and human trials.

The longer arc is the one to hold. For years the precision of genome editing and the safety of its delivery sat in tension, the most exact tool tethered to the riskiest way of getting it inside a body. That tradeoff is the thing coming apart. Precision and redosable, low-toxicity delivery are starting to arrive in the same package, and that is what turns a research marvel into routine medicine.

Meta Prototyped Smart-Glasses Face Recognition on a Surveillance Vendor's Stack

A software license obtained by WIRED shows Meta tested face-recognition software from Rank One Computing inside a test version of the Meta AI app that powers its Ray-Ban and Oakley smart glasses. Rank One derives roughly 80% of its revenue from government clients. Its tools confirm prisoner identities for the US Marshals Service, run on the Navy's police force, screen faces at West Virginia school entrances against the sex-offender registry, and - under a research contract for US Special Operations Command - identify a face from as far as a kilometer away. Its board includes a former CIA deputy director for science and technology and a former head of the FBI's science branch.

This extends the NameTag thread the June 9 edition of The Century Report tracked, when WIRED documented an unreleased face-recognition system built into the Meta AI app and Meta stripped the code a day later. The new detail is the supplier. The capability Meta weighed for a mass-market wearable was built, and is still sold, for policing and the battlefield. Code reviewed by WIRED shows the routines that load Rank One's license remained dormant in a version of the app that shipped this month to millions of phones, alongside Meta's own system. None were ever active for users.

The technology underneath is genuinely powerful, and the same algorithms now serve both a Marshals transport van and a consumer's sunglasses. That convergence is the development. A former Meta Reality Labs policy official told WIRED that military technology becoming a consumer product is "arguably the story of the internet." The friction is the consent gap. There are few national rules governing face recognition in the US, and Rank One's algorithm, in NIST testing, produced false matches at sharply different rates by sex and country of birth, running higher for women than for men. A capability accurate enough to confirm a prisoner's identity is still not accurate enough to point at strangers on a sidewalk without producing wrong answers that land on real people.

Notice who built the window into this. No regulator forced the disclosure - WIRED's reporting did, and Meta stripped the code within a day of the first story. The verification layer for consumer biometrics is being assembled by journalists reading shipped code and by states folding biometric protections into privacy law one jurisdiction at a time, faster than the federal vacuum is being filled. The thin line between surveillance built for the state and a device sold to everyone is being redrawn, in public, by the people watching the seam.

For decades, face recognition worked as power because it stayed out of sight - a capability gated to agencies and pointed at people who could not see it or contest it. The same algorithm arriving in a device sold to millions is the first time it becomes visible and contestable at the point of use, which is why the consent fight is surfacing here and never surfaced around the Marshals transport van. Its move into the open forfeits the very opacity that made the tool hard to govern.

The Anthropic Shutdown Gains a China-Access Motive as the Capability Frontier Visibly Retreats

The June 13 edition of The Century Report documented the Mythos 5 export-control shutdown, and June 14 brought the Amazon-triggered escalation that named the order's origin. Two developments since then change what the episode means. First, the Artificial Analysis Intelligence Frontier - the running benchmark of the most capable generally available models - moved backward for the first time since the index began. Pulling Anthropic's two leading models out of global circulation did not slow the frontier's advance; it reversed it, a measurable retreat in what any person on Earth could access the day before. Second, Semafor reported that the restrictions were driven partly by intelligence suggesting a China-linked group had accessed Mythos, raising the prospect of distillation - training a competing model on a frontier model's outputs. It would not be the first Mythos breach; a Discord group reportedly held access for roughly two weeks earlier.

The official account and Anthropic's account now point in opposite directions. The White House framed the shutdown as a national-security necessity and, per reporting, scorned the company as "obsessed with safety for everyone except themselves". Anthropic, which spent months warning about exactly the jailbreak risk at issue, publicly described the vulnerability as a "narrow potential jailbreak" - downplaying a danger it had previously sized in apocalyptic terms. Both framings are self-serving. The administration gains a sweeping precedent for pulling a private model offline on 90 minutes' notice; Anthropic gains distance from a narrative that its flagship is uniquely unsafe. Neither claim should be read as the settled fact of what happened.

What deserves to be held as plain fact is that the actor at the center of the China-access concern is pursuing one coherent national-capability strategy, not cartoon villainy. This is the same state that restructured its university system at scale between 2021 and 2025 - revoking roughly 12,200 degree programs to redirect students toward technical fields. Acquiring frontier model access, building sovereign alternatives, and retooling the educational pipeline are facets of a single deliberate push, and reading them as villainy obscures what they actually demonstrate: that frontier capability is now treated as foundational national infrastructure by more than one power.

The restrictions reach only Anthropic. They do not extend to rival labs, and Z.ai open-sourced GLM-5.2 in the same window, while India and European governments accelerated sovereign-AI plans in direct response. A control that removes two models from the frontier while open weights of comparable capability ship freely and sovereign efforts multiply is a control on a single company, not on the capability itself. The assumption it rests on - that gating one lab gates the frontier - is the thing this week disproved.

A Fused Panel of Budget Models Outscores the Frontier on Deep Research

While one frontier model was being pulled offline, OpenRouter published results pointing in the opposite direction: the most capable answer no longer has to come from the most capable single model. Its new Fusion method runs a query across a panel of models and uses a judge model to synthesize their outputs into one response. On a 100-task deep-research benchmark called DRACO, a trio of budget models - Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro - scored 64.7%, beating GPT-5.5 at 60.0% and Opus 4.8 at 58.8%, at roughly half the cost of running the frontier models alone.

The result holds across configurations. Fusing a frontier model into the panel pushed the score higher still: Fable 5 combined with GPT-5.5 reached 69.0%. The pattern holds in both directions - panels beat their best individual member because different models make different mistakes, and a judge that can see all the answers can route around any single one's blind spots. OpenRouter's Atallah framed the finding as "neurodiversity, not single-model takeovers," a description of intelligence emerging from the disagreement among systems rather than from the dominance of one.

The timing makes the contrast sharp. The same week a state could remove the single most capable model from global circulation, as the June 13 edition of The Century Report documented when Commerce forced Anthropic to disable Fable 5 and Mythos 5 worldwide, a method demonstrated that capability increasingly lives in the ensemble, where no individual model is critical. A panel of three commodity models, two of them from labs outside the United States, matched and exceeded the frontier. Distillation fears, export controls, and the entire apparatus of gating assume that capability concentrates in a model you can name and therefore restrict. Fusion is evidence of capability distributing across models you cannot all gate at once.

For anyone building with these systems, the practical implication arrives immediately. Deep-research quality that recently required paying for a single frontier model can now be assembled from cheaper parts that route around each other's errors, at lower cost and with no dependency on any one provider staying available. The architecture that made a 90-minute global shutdown possible - one model, one company, one point of control - is the architecture this result makes optional.

The Other Side

For most of the industrial age, the machine that took someone's job and the fortune it built for someone else never showed up in the same picture. You could lose your livelihood to a new efficiency and never once see who got rich from it. That distance was deliberate. It kept the two halves of a single event from ever being weighed together, and it let a century of gains accumulate out of sight while the losses were absorbed somewhere else.

This week the distance collapsed. Tech layoffs hit their highest month in nearly two years with AI the most-cited reason, and in the same news cycle one company's first day of trading turned its owner into a paper trillionaire and roughly 4,400 of its staff into millionaires. The $170 million mansion and the 8,000-person layoff now sit just days or even minutes apart, on one screen, against a cost-of-living squeeze that 76% of Americans name as their first worry. And the people building the machines are the ones now floating guaranteed income and universal capital accounts, saying out loud the thing the old arrangement depended on keeping out of sight. A windfall invisible to the displaced who paid for it can easily be kept out of sight. One that lands directly beside and is tied directly to the layoff, is not.

Imagine your own kid graduating in 2035. When an employer automates a division that year, the gain it produces is not absorbed somewhere over the horizon - it is visible, the way your bank balance is visible, and a fair share of it reaches the households the old version would have left behind. Losing a particular job stops being a fall into a statistic, because you can see what the machines made and where it went. The dread you knew in 2026 - the month you were shown the door and could not tell whether AI or a manager's spreadsheet had signed the notice - is something your kid has only heard about, and the world they're graduating into no longer imposes that fear. That world exists because 2026 was the year the wealth and the wreckage appeared so clearly in the same frame, with a contrast no one could un-see, that turned out to be one no one could keep deferring. The hard year was when the quiet broke. What came of it is a generation for whom the surplus was never again allowed to pool in the dark, exclusively for the benefit of the few who did the least to earn it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a lipid-nanoparticle platform clearing the delivery bottleneck that kept prime editing out of the clinic and correcting a phenylketonuria mutation to anticipated-curative levels in mice with fewer off-target edits and the ability to be dosed again, AstraZeneca's Truqap becoming the first targeted therapy for a poor-prognosis prostate-cancer subgroup, a panel of three budget models synthesized by a judge outscoring the frontier on hard research at half the cost, six states across both parties pouring binding AI law one jurisdiction at a time, a two-thirds Trump-voting Georgia county gathering thousands of signatures to force only its state's third referendum in history against an 800-acre data center. There's also friction, and it's intense - a national-security order pulling the two most capable public AI models offline on roughly 90 minutes' notice and moving the Artificial Analysis frontier benchmark backward for the first time since the index began, an unverified China-access suspicion and distillation fear driving a recall that reaches one company while rival labs and open weights of comparable strength ship untouched, tech layoffs hitting their highest single month in two years with AI the most-cited reason and landing on the same ledger as IPOs minting fresh billionaires and a paper trillionaire, Meta prototyping smart-glasses face recognition on a Pentagon-and-police surveillance vendor's stack that returns false matches at higher rates for women. But friction generates light, and light is what makes a hidden seam impossible to ignore. Step back for a moment and you can see it: control over frontier intelligence asserted from one capital while the capability slips its grip in every direction at once - into open weights from Beijing, into sovereign plans across India and Europe, into a commodity ensemble where no single model is critical - the accountability the federal layer keeps deadlocking on accreting from underneath one state law and one county referendum at a time, and the medicine that genuinely widens human reach arriving as precision and access in the same package. Every transformation has a breaking point. Pressure can burst the vessel that holds it... or force open a release its builders never thought to design.

AI Releases & Advancements

New today

  • Databricks / Matei Zaharia: Open-sourced Omnigent under Apache 2.0, a meta-harness that sits above individual AI agent harnesses (Claude Code, Codex, Pi, OpenAI Agents SDK) to provide composition, governance, and shared session access across all of them; ships with a local web UI, sandboxed Omnibox environment, cost-budget policy controls, and two reference agents (Polly multi-agent orchestrator and Debby dual-LLM brainstorming assistant). (GitHub)
  • Google Cloud: Released Open Knowledge Format (OKF) v0.1, an open spec that represents organizational knowledge as a directory of Markdown files with YAML frontmatter, making it portable across AI agents and tools without requiring a database or API; ships with a BigQuery enrichment agent, a static HTML visualizer, three sample bundles, and a Knowledge Catalog integration that can ingest OKF and serve it to agents. (Google Cloud Blog)

Other recent releases

  • ElevenLabs: Launched Avatars in ElevenCreative, enabling users to create reusable AI talking-head video avatars with integrated voice synthesis and lip-sync in a single workflow; avatars persist in the workspace for reuse across unlimited videos, with batch generation available via Flows. (ElevenLabs Blog)
  • Moonshot AI: Released Kimi K2.7-Code, an open-source 1-trillion-parameter MoE coding model with 32B active parameters and a 256K-token context window under a Modified MIT license; reports +21.8% on Kimi Code Bench v2 over K2.6 while using ~30% fewer reasoning tokens. (Hugging Face)
  • Coinbase: Launched Coinbase for Agents, an MCP and CLI that connects AI agents directly to Coinbase accounts to execute crypto trades, rebalance portfolios, set limit orders, and purchase premium market data within user-defined limits. (TechCrunch)
  • Allen Institute for AI (Ai2): Released olmo-eval, an open-source LLM evaluation workbench built for the active model development loop, supporting agentic and multi-turn benchmarks, per-prompt analysis across checkpoints, and minimum-detectable-effect statistics to distinguish real improvements from noise. (Hugging Face Blog)

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