Sign in Subscribe
Century Report

Washington Builds a Model-Release Gate - TCR 06/27/26

Washington now clears GPT-5.6 and Anthropic's Mythos customer by customer, while a quieter architecture of accountability builds from below.

Ben Linford

Ben Linford

— 19 min read
Four-panel infographic: a federal gate clearing GPT-5.6 and Mythos 5 customer by customer; renewables hit 30% US power; pig-kidney xenografts; a California job tracker and NYT suit.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The keys to the most capable models are being handed around behind closed doors over the past several days. OpenAI moved GPT-5.6 into a preview the administration clears customer by customer, naming roughly twenty organizations so far, while a Commerce letter restored Anthropic's Mythos 5 to about a hundred vetted cyber-defenders and left the public-facing version in limbo. What began two weeks ago as a one-off export action has set into a repeatable approval lane governing which models reach anyone at all. No statute authorizes it. Appointees decide who is in and who is out.

That gate arrives as Anthropic makes the boldest version of the case for trusting it: that concentrating capital, compute, and political influence in its own hands is the price of keeping AI safe. The claim is coherent on its believers' terms, and it happens to justify the military contracts and near-trillion-dollar valuation the company is already pursuing. When the most powerful actors in a field converge on the idea that the public should trust them to hold the keys, that convergence is evidence of shared interest, not arrived-at truth.

The throttle reaches one thing and misses the larger one. It slows how fast labs ship, not how fast they train, so the distance between the internal frontier and the public one only widens. The capability itself does not stay fenced, reproducing across open weights and rival labs by the quarter, the way the Vesuvius team read a sealed Roman scroll end to end and published the data and code for anyone to use.

Underneath the consolidation, a quieter architecture is being built from below. California stood up a public job-loss tracker the federal apparatus never has, turning displacement from disputed anecdote into a contestable monthly ledger. The New York Times pushed its copyright fight up the stack to the compute itself. A published transgene recipe narrowed pig-kidney xenografts toward a durable organ supply, with the method in the open record rather than locked in one vault.

Two settlements are being written at once: one by directive, customer by customer, and one in courtrooms, dashboards, and open repositories, accreting from the jurisdictions and institutions closest to the people living with the result.

The 20-Minute Deep Dive

Washington's Model Gate Hardens Into a Standing Approval Process

Two weeks after the Commerce Department pulled Anthropic's Mythos and Fable 5 offline, the most capable models from both leading labs are moving again, though only through a door the government now holds. OpenAI unveiled its GPT-5.6 suite, named Sol, Terra, and Luna, into a limited preview the administration approves customer by customer, naming roughly 20 organizations so far. Hours later, a Commerce letter dated June 26 restored Anthropic's Mythos 5 to around 100 vetted cyber-defenders and infrastructure providers, while the public-facing Fable 5 stayed in limbo with no timeline.

What the June 26 edition of The Century Report documented as a gate extended to a second lab has set into a repeatable process governing which models reach the public. OpenAI said it worked with the administration ahead of launch and called the review a short-term step, adding it does not believe "this kind of government access process should become the long-term default". That is the framing of the company being throttled. Read for what the gate's authors stand to gain, the beneficiary is the apparatus deciding access. Representative Lori Trahan named the mechanism plainly: "No law. No process. No oversight. Just appointees in Washington deciding who's in and who's out."

The throttle reaches one thing and misses the larger one. It slows how fast labs ship, not how fast they train, so the distance between the internal frontier and the public one widens from here. And the capability being fenced does not stay fenced. Stanford's Alex Stamos, reviewing Amazon's own analysis of Fable, said he found no risks absent from other publicly available models, including those made in China, and called the action "about the dumbest thing they could possibly do" in a race the administration says it wants to win. A gate can hold a named artifact. It cannot hold the knowledge of how to build the capability.

Sympathy for the throttled lab should not read as a clean bill of health: Anthropic's own training-data practices remain the subject of active copyright-infringement claims. And as TechCrunch noted, the contest has stopped being one lab against another, since OpenAI now sits in the identical position Anthropic occupied, with the same disaster waiting if a haphazard approval process chills the whole field. The rules of release for the AI era are being written by directive, customer by customer, ahead of any framework that could make them legible.

The framework's absence is where conventional coverage stops, and the same facts read forward show what is forming without it. A directive can set the timeline on a named model, but the capability it fences reproduces across open weights and rival labs by the quarter, including the publicly available models Stamos notes are already made in China. The settlement that ends up governing the era is being built in that open record, while the one issued customer by customer governs only who stands in line.

Anthropic Frames Its Own Dominance as a Safety Feature

A WIRED account of Anthropic's internal logic surfaces the central claim a frontier lab makes about itself, and it is worth holding at arm's length as exactly that, a claim. Former employees describe a company whose leaders and staff refer to themselves internally as "the good guys," and that treats the accumulation of power, in capital, compute, research talent, and political influence, as the price of fulfilling its mission to steward transformative AI safely. The company was recently valued near $1 trillion and courts customers including the US military.

The claim is internally consistent, and the piece lets its believers state it fully. Helen Toner of Georgetown offers the cleanest version: a forest full of both treasure and monsters that everyone is rushing into anyway, where Anthropic's strategy is to venture farthest while investing heavily in taming the monsters. "They're very straightforward about this," Toner says. "It's just a weird enough strategy that people have a hard time hearing it." That is the most generous reading, and it deserves to be on the page.

It also deserves the standard scrutiny any power-actor's account of its own motives draws. When a company argues that concentrating capability in its hands is itself a safety feature, the argument happens to justify the capital, the compute, and the military contracts the company is already pursuing. UCLA's Shazeda Ahmed names the gap directly, noting that organizations like Anthropic tend to struggle with a lack of pluralism, the internal diversity of thought that keeps a mission honest. A high-trust culture where employees believe leadership tells them the truth is an asset, and it is also the condition under which a single worldview goes unchallenged.

The timing sharpens the point. The same lab whose model the government just restored to a hand-picked hundred organizations is the one arguing that only its stewardship keeps the technology safe - a stewardship claim that lands weeks after the June 22 edition of The Century Report reported the governing framework behind that access being named "ad hoc, personalized, opaque, possibly lawless" by a contracting law scholar with no published criteria for future disablements. When the most powerful actors in a field converge on the framing that the public should trust them to hold the keys, that convergence is evidence of shared interest, not arrived-at truth. The wonder in what these models can do is real and reproducing across open weights and rival labs by the quarter. The claim that it is safe only under one company's control is the part that tracks the company's position, and it should be read alongside that position, not in place of it.

The Copyright Fight Climbs the Stack to the Machine Itself

The New York Times became the first major news publisher to sue OpenAI in 2023, alleging ChatGPT was trained on its articles, reproduced them near-verbatim, and positioned itself as a substitute for a subscription. The newspaper has now amended that complaint, and the change in target is the development. Where the original filing treated Microsoft's supercomputing as generic cloud service, the new complaint alleges the machine was purpose-built to infringe, built, in the paper's words, for "the purpose of using essentially the whole Internet - curated to disproportionately feature Times Works - to train the most capable LLM in history".

These are allegations, and they read as one side's characterization of events. The Times alleges Microsoft both helped select which works were ingested and supplied the means to seize them without permission, weighting high-quality journalism most heavily so that level of writing could be confidently mimicked. It further alleges Microsoft's deployment of Times-trained models across its product line helped lift its market capitalization by a trillion dollars in a single year. The amendment follows a Supreme Court ruling against Sony that the paper's lawyers evidently read as opening a path to name the infrastructure builder, not just the model maker, as an instrument of the claimed harm.

What makes the move worth tracking is where it points liability. For three years the AI training-data disputes have aimed at the lab and the dataset. This complaint aims at the compute, naming the physical buildout itself as the alleged weapon. The strongest evidence the Times cites comes from discovery, including a large chunk of users' ChatGPT sessions, where users seeking to skirt paywalls coaxed out significant article excerpts by asking for "the next paragraph," and in other cases models reproduced several paragraphs unprompted.

OpenAI sits at the center of two distinct pressures in the same week. The copyright claim climbs toward its infrastructure partner, while the same company saw its newest model throttled into a government-gated preview, with access cleared customer by customer. One entity is being squeezed from a courtroom and a federal access desk at once - less a flat villain than a firm caught between the legal architecture of the old order and the access architecture of the new one. The contest underway is over which assumptions about ownership survive when intelligence is trained on the open record of human work, and the answer is being argued building by building, filing by filing, rather than settled by anyone in advance.

California Builds the Instrument Washington Hasn't

California stood up a measurement tool that the federal statistical apparatus has not, and the gap it fills is increasingly necessary. The California AI-Unemployment Tracker links several measures of an occupation's exposure to AI against the state's monthly unemployment-insurance claims, watching for the signature of labor disruption as it surfaces rather than waiting for it to register in annual surveys. Built as a partnership between the state's Employment Development Department and the California Policy Lab at the University of California, it is described as an "early warning system" - a public dashboard a state government is willing to publish and stand behind.

For two years the displacement debate has run on disputed anecdote. Oracle attributed 21,000 cuts to AI in an SEC filing; as The Century Report covered on June 21, a Gallup poll of laid-off workers found just 1% blamed AI while remote arrangements left far more workers vulnerable; New York Fed research traced graduate unemployment largely to remote work rather than automation. Each data point arrived contested, because no one held a continuous, public instrument tying exposure to actual claims. That is the thing this tracker is: the apparatus that lets the question of how much disruption AI is causing be answered with evidence instead of assertion.

The instrument is also contestable by design, which is its quiet strength. A monthly dataset built on real claims can be checked, argued with, and corrected as the methodology matures. It moves the abstraction - "AI is taking jobs," "AI isn't taking jobs" - onto a public ledger where the claim either shows up in the numbers or doesn't. The federal layer has been notably absent here, leaving external journalists and academic economists to do the catching; a state government publishing its own continuous read is a different order of accountability, accreting from the jurisdiction closest to the workers it counts.

Read forward, that visibility is the precondition for everything downstream. A displacement that registers only after the fact, in a once-a-year revision, cannot be answered in time to matter. A windfall or a loss that can be seen month by month can be measured, contested, and built around - which is exactly what the wave of worker-transition funds, statutory-voice proposals, and redistribution blueprints will need to aim at something real rather than at a feeling. The same compression that is reorganizing work is now generating the dataset by which the reorganization can be tracked and governed, and the instrument arriving alongside the disruption is what keeps the transition legible to the people living through it.

Human Transgenes Push Gene-Edited Pig Kidneys Toward Durable Xenografts

The bottleneck in xenotransplantation has never been whether a pig kidney can be made compatible with a primate. It has been how long the graft survives once the immune system finds it. A study in Nature Communications moves that question forward by testing which precise combination of human genes added to the donor pig buys the most time.

The starting point is the triple-carbohydrate knockout pig, an animal engineered to lack the three major sugar antigens that trigger fast rejection. The research team transplanted 3KO kidneys carrying four different sets of added human genes into nonhuman primates and read out the immune response at the level of which genes switched on inside the graft. Adding human transgenes sharply reduced the transcripts tied to early immune activation, and survival lengthened with them. The clearest signal came from two anti-inflammatory genes, TNFAIP3 and HMOX1: grafts carrying them showed better survival, far less infiltration by T cells and myeloid cells, and lower expression of the gene sets that mark rejection.

What that converts is an open-ended engineering problem into a narrowing search. For years the field knew that human genes had to be layered onto the knockout pig to manage protein-level incompatibility, but the optimal recipe was undefined, which left every program guessing at a different combination. Identifying TNFAIP3 and HMOX1 as the pair that quiets early inflammation gives the next generation of donor animals a specific target rather than a hypothesis. The distance between a rejection-limited demonstration and a defined, reproducible organ source shrinks each time a guess becomes a result.

The honest caveats sit on the same page. This is demonstrated capability in primates, not deployed therapy in people, and the timeline to clinical use runs through regulatory review, prospective human trials, and manufacturing of consistent donor lines. The study was supported by eGenesis, the company developing the transgenic-pig technology, and several authors are its employees, with patents filed on the methods described. That funding relationship is worth holding alongside the data, the way any industry-run readout is, even as the immune-transcript findings are concrete and measurable rather than promotional.

Read against the waiting list, the trajectory is what counts. Tens of thousands of people sit on kidney transplant lists against a human-donor supply that has never come close to meeting demand, a scarcity treated for decades as a fixed condition of the field. A defined transgene recipe that keeps a pig kidney working longer is one of the inputs that turns that scarcity into an engineering schedule, where the constraint is how fast the work gets done rather than whether enough organs exist at all.

The Other Side

For two centuries, the worth of writing has run through a single channel: control over who gets to copy it. That made sense when copying was the expensive part. The New York Times built its case on that assumption, and recently, it tried to extend the assumption up the stack, alleging Microsoft built a supercomputer for the purpose of training on the Times's work and weighting its articles most heavily.

The same filing shows how far that control has already slipped. To name a weapon, the paper had to point at one building - a specific machine in a specific data center - because the thing it actually wants to halt, training on the open record of human writing, does not live in any one building. It reproduces across open weights and rival labs by the quarter. A court can rule on the supercomputer. It cannot rule on the method.

Hold that against another recent major breakthrough: a team read a Roman scroll sealed since 79 AD without ever unrolling it, and published the data and the code for anyone to use. Two thousand years of silence entered the open record as a gift, owned by no one.

Imagine yourself in 2034, writing something because you have something to say. The market that once priced your hours against a machine's, that made you defend every copy just to recover the cost of the work, stopped being the thing your evening depended on. The gains from this whole buildout were eventually made to reach people, and a floor arrived under you before the old one gave way. The intelligence helping you was trained on everything humans ever wrote, held in common. The hard year was when the question was still 'who owns the corpus,' argued building by building, filing by filing. What came of it is that the question stopped having to be asked, because the answer that finally scaled was all of it, for everyone.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a Vesuvius team reading a Roman scroll sealed since 79 AD from beginning to end without ever opening it and releasing the data and code for anyone to use, California standing up a public job-loss tracker that turns displacement from disputed anecdote into a monthly ledger workers can check, a defined transgene recipe pairing TNFAIP3 and HMOX1 to lengthen how long a gene-edited pig kidney survives in a primate and narrow the organ shortage into an engineering schedule, US renewables reaching 30% of the country's electricity with utility-scale solar passing wind for the first time and coal output down 11.6%, and GPT-5.6 arriving as the next step on the public frontier. There's also friction, and it's intense - a federal pre-release gate hardening into a standing approval lane where appointees clear the most capable models customer by customer with no statute behind them, a near-trillion-dollar lab arguing that concentrating capital, compute, and political influence in its own hands is itself the safety feature, the New York Times pushing its copyright fight up the stack to allege a supercomputer was purpose-built to train on its work, and Tesla settling the first known Full Self-Driving pedestrian-death suit as a probe of 3.2 million vehicles opens. But friction generates edges, and an edge is where one settlement separates cleanly from the other. Step back for a moment and you can see it: a directive holding the named models behind a hand-picked list while the capability itself keeps reappearing across open weights and rival labs by the quarter, and a quieter architecture built from below - a state's dashboard, a publisher's filing aimed at the compute, an openly published scroll, a transgene recipe in the public record - rising from the states, courtrooms, and research groups nearest to the workers, readers, and patients who live with the changes. Every transformation has a breaking point. Sunlight can scorch the ground it shines on... or power a grid no single plant was ever large enough to carry.

AI Releases & Advancements

New today

  • 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)

Other recent releases

  • 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)
  • Google DeepMind: Added computer use as a built-in tool in Gemini 3.5 Flash, enabling the model to see, reason, and take action across browser, mobile, and desktop environments for long-horizon enterprise automation; includes adversarial training against prompt injection and two optional enterprise safeguard systems for sensitive-action confirmation and injection detection. (DeepMind Blog)
  • Alibaba Qwen: Released Qwen-AgentWorld-35B-A3B, a 35B MoE / 3B active-parameter Language World Model trained to simulate seven agentic environments (MCP, Search, Terminal, SWE, Android, Web, OS) via chain-of-thought reasoning over 10M+ real-world interaction trajectories; Apache 2.0, 256K context, compatible with vLLM and SGLang; accompanied by AgentWorldBench, a seven-domain evaluation benchmark. (Qwen Blog)
  • NVIDIA: Released NeMo AutoModel as an open library that wraps Hugging Face Transformers v5 with Expert Parallelism, DeepEP fused all-to-all dispatch, and TransformerEngine kernels for MoE fine-tuning, delivering 3.4–3.7× higher training throughput and 29–32% lower GPU memory usage than native Transformers via a single import-line change. (Hugging Face Blog)
  • Mozilla: Launched the MDN MCP Server, an official Model Context Protocol server for MDN Web Docs enabling AI coding assistants and agents to query authoritative web development documentation programmatically. (MDN Blog)
  • Microsoft Research / Centre for Population Genomics / Broad Institute: Open-sourced Talos, an automated iterative genomic reanalysis tool that recovered 90% of rare-disease diagnoses while surfacing only 1.3 candidate variants per patient for expert review; deployed across nearly 5,000 undiagnosed patients, yielding 241 new diagnoses with an average 32-day lag between new supporting evidence and diagnosis. (Microsoft Research 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