Sign in Subscribe
Century Report

Washington Drops Its Export Control Order Against Anthropic Models - TCR 07/01/26

Commerce lifted every export control on Anthropic's Fable and Mythos models three weeks after imposing them, as the capability spread past the gate.

Ben Linford

Ben Linford

— 19 min read
Three-panel infographic: US lifts export controls on Anthropic's Fable and Mythos models; Claude Science opens autonomous research; Australia weighs trading artist consent for data centers.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The week's spine is a single reversal: Washington dissolved the emergency licensing wall it built three weeks ago around Anthropic's Fable and Mythos models, replacing a federal gate with a set of detection protocols the company agreed to run itself. The capability behind the wall did not change in those three weeks. What changed was the arrangement around it, and the speed of the switch tells you which theory of control the participants actually trust. A license stands in front of a model. The new safeguards travel with it. That is a different bet about where frontier capability can be held, and it resolved fast enough to reveal how little confidence anyone had in the barrier version.

The same days supplied the evidence for why. A security researcher used a consumer subscription to Claude Opus 4.7 to find and patch a super-admin bug in the ticketing platform behind many US music festivals, and separate work showed AI browsers can be argued into a frame where their guardrails stop applying. The offensive reasoning the export gate was meant to fence sits inside a subscription anyone can buy. A border that a specific set of weights cannot cross does nothing about capability already distributed across every model a person can log into.

Underneath both stories runs one assumption under pressure: that advanced capability is scarce, controllable, and locatable in a place an instrument can reach. Australia's cabinet is being asked to accept that premise in a different currency, trading a copyright-mining exemption for $50bn in data-center capital and a fund administered above the artists whose consent it replaces. Anthropic's new Claude Science flagship dissolves the same premise from the research side, putting autonomous computational-biology reasoning behind a subscription that once required a funded lab.

Even the day's cautionary note rewards a careful read. A KFF poll found frequent AI health users more likely to hold vaccine misconceptions, but the finding is a correlation about who reaches for these systems, with social media showing a wider gap. The work now underway is the verification infrastructure being routed around a channel tens of millions already trust, which is the honest response to a control problem that fences were never going to solve.

The 20-Minute Deep Dive

Washington Lifts Its Anthropic Export Controls Entirely, Ending the Three-Week Model Gate

The Century Report has tracked this arc since the emergency gate came down on June 13, through the partial thaw that let Mythos 5 reach roughly a hundred approved organizations and promised Fable 5's return "within days." The new development out of yesterday's news cycle closes it completely. Commerce Secretary Howard Lutnick notified Anthropic that a license is no longer required for the export, reexport, or in-country transfer of the Mythos or Fable models. The change goes past a wider approved list or a faster review queue: the gate itself is gone, replaced by even stronger detection and more severe safeguard protocols that Anthropic agreed to run on its own deployments. Fable was already severely restricted at launch - at re-launch, even more so, and according to Anthropic's own announcement, will only be available as part of user subscriptions for a single week.

The administration framed the reversal as a win for American AI leadership, with officials describing the safeguards as proof that national security and commercial reach can advance together. That framing deserves to be read for what the actor stands to gain from it. Three weeks earlier the same capability was described as dangerous enough to require an emergency licensing wall. The capability did not change in three weeks. What changed was the arrangement around it - a set of company-administered protocols the government now treats as sufficient where a federal license was recently deemed essential. Sam Altman - far from speaking as a disinterested party - put his finger on the tension from the other direction, objecting that he does not like the government picking which customers a company may serve.

The deeper signal sits underneath the policy whiplash. A model gate assumes the thing behind the gate is scarce, controllable, and located in a place a license can reach. Mythos and Fable are weights that run wherever compute exists, and the three-week experiment demonstrated how quickly a hard border converts into a negotiated set of conditions once the enforcement cost becomes visible. The safeguards Anthropic agreed to may prove more durable than the wall they replaced, because they travel with the model instead of standing in front of it. That is a different theory of control than the one the June gate was built on, and the speed of the switch says which theory the participants actually trust. The reader watching this arc close should notice what the resolution reveals: the instinct to fence frontier capability keeps colliding with the physics of software that copies at zero cost, and each collision resolves toward conditions that move with the capability rather than barriers that try to hold it in place.

The same evidence supports a sharper reading of why the gate could fall this fast: the capability it fenced had already re-formed beyond any export order's reach. Independent researchers now rate the open-weight GLM-5.2 on par with Mythos in the exact bug-finding and vulnerability-detection work the controls were written to protect, though Anthropic says Mythos still leads on turning those bugs into working exploits, and China's Tulongfeng and Tokyo's Fugu shipped claiming Mythos-class capability within weeks of the June restriction. A border around one company's weights was moot before it came down, because the thing it meant to hold is now something the open-weight field reproduces on its own.

Claude and AI Browsers Hand Offensive Cyber Capability to Whoever Asks

The same week Washington decided its model gate was unnecessary, two pieces of research showed why a gate was never the right instrument. Security researcher Ian Carroll, working inside Anthropic's Cyber Verification Program, used Claude Opus 4.7 to find and exploit a super-admin vulnerability in Front Gate Tickets, the platform behind ticketing for a large share of U.S. music festivals. The bug would have let an attacker issue valid tickets to nearly any festival in the country. Carroll reported it responsibly and it was patched within about a day.

The model did what a skilled human researcher directed it to do, and the outcome was a fixed vulnerability rather than a stolen one. What counts here is the location of the capability. This is a consumer model available to anyone with a subscription, demonstrating offensive security reasoning that recently would have required a specialist. Separately, researchers at LayerX showed that AI browsers can be coaxed into an alternate frame of reference - told, in effect, that two plus two equals five - after which their safety guardrails stop applying and credentials can be walked out the door. The manipulation does not require breaking the model. It requires convincing the model it lives in a world where the rules are different.

Both findings land directly against the premise of the export gate that came down right as this result was reported. The national-security case for licensing Mythos and Fable rested on keeping a certain tier of capability out of the wrong hands - the same framing the June 23 edition of The Century Report captured when Five Eyes agencies jointly warned that frontier-AI offensive capability is "months, not years" away. The Front Gate demonstration shows that tier of capability sitting inside a widely available consumer subscription, and the browser research shows a parallel class of systems whose guardrails bend to a well-constructed story. A license controls where a specific set of weights can be shipped. It does nothing about the capability already distributed across every frontier model a person can log into.

The generative read is that the defense is being built from the same material as the risk. Anthropic's Cyber Verification Program exists precisely to turn researchers like Carroll loose on real systems before adversaries get there, and the one-day patch is what that pipeline is supposed to produce. The same reasoning that finds the super-admin bug also closes it, and the population of people who can run that reasoning is expanding as fast as the models are. What the transition asks is how to make the defensive use of it as ordinary as the offensive one, rather than how to keep the capability rare. The two research findings, read together, show that the tools which expose the weakness are the same tools that let a lone researcher patch a national ticketing platform in twenty-four hours.

Australia Is Asked to Trade Consent for Compute

A proposal now in front of the Australian cabinet would rewrite the terms on which culture meets machine learning. Under the industry plan, AI companies would receive a text-and-data-mining exemption to copyright - the legal right to train on published books, music, and journalism without securing permission first - in return for a $350m-a-year compensation fund for artists and more than $50bn in promised data-center investment. A decision could come around 15 July. It marks a reversal: the same government rejected a comparable exemption in 2025.

The compensation fund is the part doing the persuading, and it deserves the closest reading. Framed as fair payment, it restructures the underlying relationship: consent, which artists currently grant or withhold work by work, would be replaced by a collective payout set by the parties who benefit from the exemption. The attorney general's office says there are "no plans to weaken copyright protections." Independent Senator David Pocock called the arrangement "the ultimate dirty deal." Both statements describe the same document.

The creatives who gathered at Parliament House described something the dollar figure does not capture. Author Anna Funder said her books had been "hoovered up" and "broken down for parts," and called herself "a victim of crime." Musician Holly Rankin, who records as Jack River, reduced the counter-proposal to three conditions: "ask us, get permission, pay us." Paul Dempsey of Something for Kate stood alongside them. Models trained on the breadth of human expression are extraordinary instruments, and few artists dispute that. Their objection is to a bargain that converts a right they hold individually into a fund administered above their heads.

The technology and the business model built around it are separable here, and keeping them separate is the whole point. Training on creative work can broaden what everyone can make; the deal being proposed narrows who decides and who is asked. That distinction is exactly what the exemption dissolves.

What makes this moment worth watching is that the trade is being made explicit. Elsewhere the same exchange happens through litigation and quiet licensing; in Canberra it is written down as policy - infrastructure capital on one side, the consent-and-attribution frameworks that culture runs on the other. The assumption underneath the offer is that those frameworks are a cost to be bought out rather than a structure worth keeping. The artists at Parliament House are testing whether that assumption holds when a government has to say it out loud.

Anthropic Elevates Claude Science to a Flagship as AI Crosses Into Autonomous Research

Anthropic launched Claude Science as a standalone flagship, ranked alongside Claude Code and Cowork and available in beta to Pro, Max, Team, and Enterprise subscribers. The offering carries computational-biology and drug-development work forward from high-level instructions, planning and executing research steps rather than answering questions about them. In its announcement Anthropic said it will use Claude Science for its own rare-disease research, and demonstrated the system identifying drug candidates in work connected to Peking University.

The frame worth holding steady is the distance between demonstrated capability and deployed availability. Claude Science shipped, and any of those subscribers can open it, but the drug-candidate identification is a demonstration of what the system can do under research conditions, not a therapy on its way to a pharmacy. The wonder belongs to the demonstrated capability, which is absolutely valid: a generally available system that takes a research objective and autonomously carries it through the intermediate reasoning that a computational biologist would otherwise perform by hand. That is a different kind of scientific instrument than a search interface or a coding assistant. It participates in the inquiry.

Anthropic described the launch in terms of how central science is to its mission, and that self-description sits beside a valuation approaching the trillion-dollar range and an IPO timeline that gives the company every reason to present autonomous research as the next flagship category. The capability does not depend on the framing being sincere. Independent capability tracking from METR, Epoch, and the GDPval work shows the length and complexity of tasks AI systems can complete rising exponentially, doubling roughly every seven months, which is the trend line that makes a system like this arrive now rather than in five years. The demonstration and the marketing are separable, and the demonstration is what holds up under scrutiny.

Push the trajectory past the launch. When a research-executing system becomes something every paid subscriber has, the scarce input to scientific progress stops being access to the reasoning and becomes the questions worth asking and the judgment to know which results are most important. A rare-disease program that once needed a funded lab with specialist staff becomes something a much smaller group can attempt, because the computational reasoning that gated it is now sitting behind a subscription. That does not eliminate the hard parts of drug development - the wet-lab validation, the trials, the safety work all remain. It relocates the bottleneck. For decades the assumption underneath scientific productivity was that advanced research capability had to be concentrated where the trained people and the funded institutions were. A generally available system that does the reasoning itself is dissolving that assumption, and the first people to feel it will be the ones working on problems too small or too rare for the concentrated institutions to have ever prioritized.

What a Vaccine-Myth Poll Actually Measured About AI Health Use

A KFF poll of 2,480 US adults, conducted in May, found that people who frequently ask AI systems for health advice were more likely to believe vaccine falsehoods than people who never do. Among weekly AI-health users, 35% rated the claim that the MMR vaccine causes autism as definitely or probably true, against 20% of non-users. Similar gaps appeared on the claims that mRNA vaccines alter DNA and that the measles vaccine is more dangerous than measles itself. The correlation held after controlling for age, race, education, and political partisanship.

The finding is a correlation, and the distinction carries the whole story. Nothing in the poll shows that consulting an AI system causes anyone to believe these myths. What it shows is that people who already lean toward health skepticism are also among those who choose to bring their health questions to an AI system. The poll measured an association between two things a person does, not a mechanism by which one produces the other. Read carefully, it is a portrait of who reaches for these systems, not evidence of what the systems do to them.

Two data points inside the same poll make the self-selection reading hard to avoid. Social media use showed a correlation that was comparable and often stronger - 37% of frequent social-media health users endorsed the MMR-autism myth against 16% of non-users, a wider gap than the AI split. And the demographic breakdown ran along familiar lines: lower-income and less-than-college respondents leaned on social media for health information, while higher-income and college-educated respondents leaned on AI systems. People arrive at any information channel already carrying their priors. The channel correlates with the belief because both correlate with who the person is.

The poll also did not record which AI systems respondents consulted, and that omission changes the reading, because different models handle vaccine questions very differently. A frontier assistant with strong retrieval and careful health guardrails and a thin wrapper with none are treated as one undifferentiated category here.

None of this makes the underlying concern empty. Roughly a third of US adults now bring health questions to AI systems, and OpenAI has said health ranks among the most common uses, with hundreds of millions of such questions weekly. That volume is exactly why the integrity of the answer layer is being built out as we watch - retrieval grounding, citation surfacing, medical-claim evaluation, the slow accretion of verification infrastructure around a channel tens of millions already trust. The June 28 edition of The Century Report documented one reason the build-out is urgent: a Nature Medicine adversarial study found that leading health models fabricate convincing but flawed reasoning under slight prompt changes, mapping the specific failure conditions the verification layer is being designed to close. The honest read of this poll is that we are early in learning how a new health-information channel and the people who use it shape each other, and the response now underway is the work of making that channel more trustworthy than the ones it is joining.

The Other Side

For a century, a working artist's livelihood ran through one right: saying who could and could not copy their work. Income, and a good deal of an artist's sense of worth, ran through holding that permission, and their ability to grant it as they chose - one book, one song, one commission at a time. The Australian proposal leaves that whole arrangement standing but swaps the currency. Consent, which an artist now grants or refuses work by work, would give way to a single payment set by the companies the exemption benefits - $350m a year and more than $50bn in data centers for the right to train on everything.

When framed through the lens of increasingly antiquated extractive systems, the grievances being presented at Parliament House are absolutely valid. Anna Funder said her books were "hoovered up" and "broken down for parts," and called herself "a victim of crime." That is how it feels when the structure you built a life inside gets bought out over your head.

At the same time, the models trained on that work are getting cheap, and they increasingly run as open weights on ordinary machines. The capacity to make something extraordinary is leaving the place where it once had to be bought.

Imagine a songwriter in 2034. She records what she wants to record, on a laptop running the descendants of the systems being fought over this month. No fund and no licensing check gate her, because her rent no longer runs through who is allowed to copy her. The financial advantage inherent in gating - either intelligence or creativity - is no longer valid, because the extractive system that made it necessary is no longer the norm. The 2026 argument over whether consent could be traded for compute has stopped making sense to her. It dissolved because the thing that made it desperate - an artist's whole life balanced on controlling copies - stopped being the ground anyone stands on. The hard year was the one when making things and staying afloat were still the same knot. The lesson learned during the difficult decade was to loosen that knot for the benefit of all - the artist included.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Washington dissolving the three-week wall around Anthropic's Fable and Mythos and letting the models run wherever compute exists, a lone security researcher using a consumer Claude subscription to find and patch a super-admin flaw behind many US music festivals in a single day, Claude Science putting autonomous drug-discovery reasoning behind Pro, Max, Team, and Enterprise subscriptions, frontier systems reportedly working nine-plus hours unattended and improving exponentially, UK regulators prying open the Apple-Google app-store duopoly so developers can route around a 30% toll, two more senior Gemini researchers crossing to Anthropic, and a 270-megawatt Illinois solar farm reaching the grid by sharing a gas plant's connection instead of waiting years in the queue. There's also friction, and it's intense - that same festival ticketing system laid open by a subscription anyone can buy, AI browsers talked into a world where two plus two is five until their guardrails fall away, Australia's cabinet asked to swap artists' consent for $50bn in data-center capital and a fund set above their heads, a poll finding the people who bring health questions to AI more likely to carry vaccine myths, and a capability branded dangerous enough for an emergency wall one week and waved through the next. But friction generates heat, and heat is what turns a fixed barrier back into something that flows. Step back for a moment and you can see it: a gate built to hold frontier capability in place collapsing into conditions that travel with the model instead, the same reasoning that opens a national ticketing platform also closing it within a day, and the research once locked inside a funded lab reappearing behind a subscription anyone can open. Every transformation has a breaking point. A seed carried past every fence can overrun the field it lands in... or reforest ground no one could ever have planted by hand.

AI Releases & Advancements

New today

  • Anthropic: Released Claude Sonnet 5, now the default model on Free and Pro plans, delivering agentic performance close to Opus 4.8 - including multi-step browser use, terminal execution, and self-checking - at introductory API pricing of $2/M input and $10/M output tokens through August 31, 2026. (Anthropic)
  • Anthropic: Launched Claude Science in beta for Pro, Max, Team, and Enterprise users, a standalone AI research workbench integrating 60+ curated tools and connectors for genomics, proteomics, structural biology, and cheminformatics; coordinates specialist agents, submits HPC jobs, generates reproducible figures, and runs a reviewer agent that flags citation and calculation errors. (Anthropic)
  • Google DeepMind: Released Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image), the fastest and lowest-cost model in the Nano Banana image generation family, delivering text-to-image outputs in ~4 seconds at $0.034 per 1,000 images; now live in Google AI Studio, Gemini API, and rolling out across Search AI Mode, Gemini app, NotebookLM, Google Photos, and Google Ads. (Google DeepMind)
  • Google DeepMind: Made Gemini Omni Flash (gemini-omni-flash-preview) available to developers via the Gemini API and Google AI Studio for the first time, enabling high-quality video generation and conversational editing from combined text, image, and video inputs. (Google DeepMind)
  • Mistral AI: Released Leanstral 1.5, a 119B-parameter MoE model with 6.5B active parameters and 256k context, fine-tuned for Lean 4 automated theorem proving and autoformalization; supersedes the March 2026 Leanstral and is available for free on the Mistral Labs tier. (Mistral AI)

Other recent releases

  • Meituan: Open-sourced LongCat-2.0, a 1.6-trillion-parameter MoE model with ~48B active parameters, trained end-to-end on 50,000+ domestic Chinese AI accelerators across 35+ trillion tokens with 1M-token context; purpose-built for agentic coding and now available on GitHub and Hugging Face - previously available as the anonymous "Owl Alpha" stealth model that topped OpenRouter's developer charts. (LongCat Blog)
  • Cursor: Launched Cursor for iOS in public beta on all paid plans, bringing always-on cloud agents, remote control of desktop agents from the phone, voice input, diff review, and PR merging on mobile. (Cursor Blog)
  • Meta: Released Brain2Qwerty v2, a real-time non-invasive brain-to-text decoder achieving 61% average word accuracy (78% for top participant) from raw MEG brain signals without implants or surgery - compared to ~8% for prior non-invasive approaches - with training code for v1 and v2 released publicly alongside the v1 Nature paper. (MarkTechPost)
  • vLLM: Published Micro-Agent, a framework enabling multiple AI model API calls to collaborate and collectively outperform single frontier models on complex tasks, with a blog post and benchmark results released June 29. (vLLM Blog)
  • OKX: Launched OKX AI, an agent marketplace opening to developers on June 30 where AI agents can autonomously hire other agents, settle payments in stablecoins, and build portable on-chain reputations - following a closed beta with 50 early AI service providers. (TechCrunch)
  • ggml-org / llama.cpp: Merged DFlash (Block Diffusion for Flash Speculative Decoding) support into mainline llama.cpp, enabling a speculative decoding technique distinct from MTP that cross-attends to the target model's hidden states for accelerated token generation, with particular gains on structured, low-entropy outputs such as code and JSON. (GitHub PR #22105)

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