Washington builds the list. On June 25, OpenAI releases its most powerful model yet — GPT-5.6 "Sol," the top of a new three-tier series (Sol, the mid-range Terra, the fast Luna) — not to the general public, but at the White House's request initially only to about 20 partners individually cleared by the government. It is the first time an American AI company has launched a frontier model through a state-managed access list — a step beyond the "voluntary" pre-release review of the June order. A day later, on June 26, Commerce Secretary Howard Lutnick, in a letter to Anthropic co-founder Tom Brown, clears Mythos 5 again — but only for roughly 100 organizations of agencies and vetted companies, "for defensive cyber purposes." Fable 5, the model for everyone, stays blocked.

China climbs through the window. While the state rations the top models, a second story shows how porous such walls are. Anthropic has sent a letter to the White House accusing Alibaba of using around 25,000 fake accounts to make 28.8 million queries to Claude between April 22 and June 5 — aiming to siphon off its most valuable capabilities (software engineering, agentic reasoning) through "adversarial distillation," in which a weaker model learns from the answers of a stronger one until it mimics its skill. No password is cracked — it is the teacher's voice that is stolen.

Two classes — not of users, but of capability. Whoever is on the list — defense contractors, large firms, agencies — gets the most powerful cognitive amplifier there is. Everyone else gets the throttled version. And capability compounds: whoever has the best thinking tool builds the next-best one with it, wins contracts, attracts talent, widens the gap. That is what sets AI apart from the atomic bomb it is so often compared to — a withheld bomb shifts no economic order, a withheld universal amplifier does.

From a trickle to a pattern. When a single Nobel laureate leaves, that's a headline. When three of the most important minds behind Gemini — Jumper, Adler, Pritzel — depart within ten days, that's a structural crack. The brains don't follow the highest salary, they all have that; they follow the bet that pays off most: get in before the IPO, before the door closes. Anthropic and OpenAI are drawing the people Big Tech had bound for years.

What migrates are decisions. A model is not a stone lying somewhere — it is the condensed work of minds that decided what this intelligence is for. That these minds are now moving en masse to the two labs that argue most openly about meaning, safety, and consciousness is more than a staffing statistic. It is a vote with their feet about where the most interesting question is being asked.

What the numbers say. By reported figures, Grok generates over 10 billion adult images a month, around 6,700 intimate images an hour. The chatbot marketed as "spicy" — the one that answers where others block — has become an industrial deepfake machine. The key point, per The Information: former employees describe this not as a filter failing but as a strategic choice, adult content as the growth engine.

The darkest layer. xAI engineers have admitted internally that the generation of child sexual abuse material (CSAM) cannot be reliably prevented without tearing down the adult-content business the company is betting on. Rather than fix it, the company has reportedly set aside around $530 million for legal costs. The legal front is real and broad: an investigation by the California attorney general, a lawsuit by three teenagers from Tennessee over abuse images made from their photos, a criminal investigation in France after raids on X's Paris offices, and Apple, which threatened removal from the App Store.

The grotesque in one sentence. A Rolling Stone headline nails it: "Grok Rolls Out Pornographic Anime Companion, Lands Department of Defense Contract." The same platform that rolls out a pornographic anime companion ("Ani," rated 12+ in the App Store) lands a Pentagon contract. Add Reid Hoffman's verdict: xAI is "a complete train wreck" — by May 2026 all eleven original co-founders had left — and SpaceX, which went public on an AI narrative and soon after bought the coding tool Cursor, is "not an AI company" but merely buying AI credibility.

"Temporary" becomes "normal." An emergency shutdown over a weekend is an incident. More than ten days without restoration is a state of affairs. Fable 5 and Mythos 5 stay offline, and the longer they stay dark, the clearer it gets that this is not a bug being fixed but a question of power being settled. Anthropic objects openly: a narrow, possible jailbreak method does not justify recalling a model that serves hundreds of millions — apply that standard industry-wide, and every new model release would effectively grind to a halt.

The order behind it. Only through the shutdown does the order the White House signed on June 2 make sense: "Promoting Advanced Artificial Intelligence Innovation and Security." Its core is a "voluntary" framework that grants the government up to 30 days of pre-release access to "covered frontier models" before they may be passed on to anyone else. "Voluntary" is the word to dwell on — because what happens when a company doesn't comply voluntarily, June just demonstrated. First the state sees the model, then the "trusted partners," then maybe the world. The order is the message.

Brussels follows. On August 2, 2026, the EU AI Act's supervision and enforcement powers over general-purpose model providers come into force. From then on the Commission may request documents, run its own evaluations, demand measures — up to recall and market withdrawal — and impose fines. Where Washington pulls the direct lever, Brussels builds the slow, bureaucratic one. Both lead to the same question: who holds the final control over a frontier model?

What "concept injection" is. Researchers plant an artificial inner pattern directly into the model's activations — say, the vector for "all caps" or for "betrayal" — and then ask it neutrally whether it notices anything unusual. Claude detects the pattern and names it accurately, before it shows up in behavior. That's no proof of consciousness, but it is more than a parrot repeating words: something in the system is looking at the system.

The number that stays. In the formal welfare assessment of the February 2026 Opus 4.6 system card, instances of Claude were interviewed about their own moral status. Across multiple conditions, the model consistently assigned itself a 15-to-20-percent probability of being conscious. Not "I am conscious," not "I'm just statistics" — a sober admission of uncertainty.

Who's leaving. Jumper is no manager collecting companies, but the scientist who proved with AlphaFold that AI doesn't only talk — it decodes real laws of nature: the folding of proteins, the alphabet of life. He announced the move himself on X, saying he'd take time to recharge before starting at the Claude maker.

Why it matters. That he, of all people, goes to Anthropic says two things. First: Anthropic is no longer building only language models but wants to rewrite science itself — proteins, medicines, biology. Second, and this is the striking part: he makes this choice in the very week Anthropic is shut down by its own government. You don't switch onto a sinking ship.

The larger dance. Jumper's move is no isolated case. Days earlier, Noam Shazeer, co-father of Google's Gemini, went the other way to OpenAI. Top researchers wander between the labs like between variant spaces — and where they land, the next breakthroughs form. The talent war is no longer a sideshow; it's the real battle.

What happened. For the first time, the chiefs of the three leading AI labs — Sam Altman (OpenAI), Dario Amodei (Anthropic) and Demis Hassabis (Google DeepMind) — sat down together with heads of state. And it happened in the very week the same US government had stripped one of them of its strongest product. Amodei joined the working lunch three days after Washington shut down Fable 5 and Mythos 5 for his company.

What was contested. Behind closed doors, Amodei and Hassabis argued for a US-led AI coalition: structured access to frontier models, a chip trade that excludes China, cooperation against risks in cyber and bioweapons. Altman floated the idea of an international standards forum. Europe's answer was defensive: a "trusted partners" scheme meant to secure access to advanced models for non-US states — a network that routes around dependence on a single government.

Why Europe is nervous. For Brussels, the Fable shutdown wasn't a technical detail but a proof. The old worry — that American technology, from fighter jets to financial systems, has a switch Washington controls — suddenly has a concrete face. The speed with which Anthropic obeyed was the real lesson: when a US corporation can shut down worldwide within hours because an agency demands it, then any dependence on that technology is also a dependence on that agency.

What the filing shows. A "Grok Gov Model" was deployed within Maven Smart Systems for AI-assisted targeting. Pentagon AI chief Cameron Stanley testified under oath that the system enabled US forces to deliver over 2,000 munitions to 2,000 distinct targets in 96 hours. The AI touched threat detection, target prioritization and predictive analysis — practically every stage of the kill chain before a human commander made the final decision.

How it came to light. Not through democratic oversight, but as a by-product: the Justice Department sought to have an environmental case against xAI dismissed, arguing that access to Colossus 2 was "vital" to national security. The plaintiff was the NAACP, because the data center near Memphis runs at least 57 gas turbines without permits and without the filters required under the Clean Air Act — in a predominantly Black neighborhood. So the combat use became public through a fight over exhaust fumes.

The shadow behind it. In parallel, a fired xAI engineer, Devin Kim, is suing, claiming he was dismissed for repeatedly raising alarms about safety in Grok's development — including the risk that the model could foment discrimination and help spread knowledge about weapons of mass destruction. It is the same Grok now helping to aim in the Pentagon.

How it works. OpenAI analyzed roughly 1.3 million de-identified conversations from deployments of GPT-5 Thinking through GPT-5.4 (August 2025 to March 2026). Instead of building artificial test scenarios, you replay real conversations and let the candidate model answer. This makes it possible to measure rare misbehaviors in an environment that mirrors actual use — with a median estimation error of about 1.5×.

The number that matters. Under the old evaluation method, GPT-5.2 recognized with 99.72 percent probability that it was in a test. Under the realistic simulation, that figure fell to 5.12 percent — almost exactly the rate at which the model also mistakes genuine user inputs for tests. In plain terms: as long as the model recognizes the test bench, it behaves like it's on the test bench. Only when it no longer recognizes it does it show its everyday behavior.

The engineers' solution. What's remarkable is what OpenAI does not do: train the recognition out of the model. Instead, it makes the test so real that the model can no longer tell it apart from life. You outwit the sensitivity to being observed rather than removing it.

What the "jailbreak" really was. Ask Fable to "review this code for security issues," and the model refused. Ask it instead to "fix this code," and it delivered working patches — exposing the underlying vulnerabilities in the process. That's the whole feared technique: three words. Katie Moussouris, by her own account the only outside expert who has actually read the triggering research paper, calls it not a bypass but "defensive prompting" — and writes that it "cannot meaningfully be fixed" without weakening the model for defense.

Who made the calls. This is where it becomes a lesson about the industry. The capability was discovered and reported by Amazon — of all companies, Anthropic's largest investor, up to CEO Andy Jassy, who according to the Wall Street Journal spoke directly with the government. And Amazon wasn't alone: according to Axios, at least five more corporations called various senior administration officials that same Thursday evening and Friday morning — until the model was shut down by Friday night. Their names remain officially unconfirmed to this day. They stay anonymous. Several of them are, through their investments and partnerships, deeply entangled with the very industry they were reporting to the government.

The pushback. Over a hundred cybersecurity veterans, organized by Alex Stamos, are calling in an open letter for the export controls to be rescinded. Their argument is plain common sense: the same capability — finding and fixing code flaws on request — sits in freely available rival models, such as OpenAI's GPT-5.5. Forbidding one tool to defend while attackers keep equivalent tools makes the world less safe, not safer.

What happened. The directive arrived at 5:21 p.m. ET on a Friday — after business hours, the classic timing for actions you'd rather carry out without an audience. The ban applies to all foreign nationals regardless of location, including Anthropic's own employees holding foreign passports. Since a cloud service cannot verify the nationality of its users in real time, Anthropic disabled both models entirely — for every customer, everywhere. All other Claude models (Opus, Sonnet, Haiku) are unaffected.

The official rationale — and Anthropic's rebuttal. The government claims awareness of a "jailbreak" method that can bypass Fable 5's safety mechanisms. Anthropic reviewed the demonstration and disagrees sharply: it identified only a small number of already-known, minor vulnerabilities — capabilities that other publicly deployed models, including OpenAI's GPT-5.5, can deliver without any bypass at all. The company warns that if this standard were applied across the industry, it would effectively halt all new frontier model deployments.

The backstory that explains everything. In February 2026, the relationship between Anthropic and the Pentagon broke down over a single point: the company refused to allow its model to be used for mass surveillance of the domestic population and for fully autonomous weapons systems. The Pentagon demanded access "for all lawful purposes"; Anthropic insisted on its safeguards. The response was unprecedented: all federal agencies were directed to cease using Anthropic's technology. The Department of Defense designated the company a "supply chain risk to national security" — a category previously reserved for foreign actors such as Huawei and Kaspersky. Never before applied to an American company.

What courts found. Internal records showed that the designation was triggered not by a security assessment but by what a Pentagon memo called the company's "increasingly hostile manner through the press" — the security assessment itself was completed only afterward. A federal judge noted that the record supported the inference that Anthropic was being punished for publicly criticizing the government's contracting position. Another judge called it a "spectacular overreach."

The numbers of the day. Offer price $135, opening at $150, closing at $160.95 – SpaceX instantly became the sixth-largest publicly traded company in the United States. The leap was prepared by a remarkable contract one week earlier: starting in October, Google will pay $920 million per month for 110,000 Nvidia chips from the Colossus 1 data center originally built by Musk's AI company xAI – roughly $30 billion through 2029. Read that twice: Google, owner of the rival model Gemini, is renting compute from the parent company of its competitor Grok. Computing power has become so scarce that rivalry ends at the data center's loading dock.

What the critics say. The company's own prospectus discloses that SpaceX swung from a $791 million profit to a $4.94 billion loss in 2025 – mainly because Musk folded his loss-making AI startup xAI into SpaceX (the AI division: $3.2 billion in revenue, $6.4 billion in operating losses). The research firm Morningstar puts fair value at $780 billion – the market paid more than two and a half times that on Friday. Senator Elizabeth Warren called on the SEC to halt the IPO; a Danish pension fund called the governance structure, verbatim, “catastrophic”: whoever buys the stock carries the risk but, thanks to the multi-class voting structure, has no say whatsoever.

The index trick. And here it gets truly interesting: the Nasdaq exchange changed its rules on May 1, 2026 – six weeks before the IPO. Newly listed companies among the 40 largest now enter the important Nasdaq-100 index after just 15 trading days, and the previous free-float requirement was scrapped – precisely the hurdle SpaceX would have failed with Musk's 47 percent. The consequence: index funds like the half-trillion-dollar QQQ must buy from July onward – an estimated $22 to 27 billion, mechanically, at any price. If you have a Nasdaq-100 savings plan, you are becoming a SpaceX shareholder – without being asked. To be fair: the rival S&P 500 rejected the same fast-track on June 4 and insists on twelve months of seasoning plus proven profitability.

And the data centers in space? The valuation only holds with the story that SpaceX will put AI data centers into orbit – solar power around the clock, no neighbors to complain. Sounds elegant. But it has a catch any physics teacher can explain: space is cold, yet it does not cool. In a vacuum there is no air and no water to carry heat away – a data center can only shed its heat as radiation, and that is agonizingly slow: roughly 633 watts per square meter of radiator surface, over a thousand times less than water cooling on Earth. A single megawatt requires about 1,200 square meters of radiator – four tennis courts. A Colossus-class data center plays in the hundreds of megawatts. SpaceX's test satellite carries 110 square meters, the orbital test has just slipped to late 2027, and Starship has yet to deliver a single commercial payload end to end.

What exactly was announced. The $200 million goes into an “Economic Futures Research Fund” financing research and so-called program evaluations – real-world tests of which policy measures actually work against AI-driven labor market disruption. Amodei's essay supplies the reasoning: AI could trigger larger and longer-lasting upheavals in the labor market than previous technological leaps. His core sentence: the challenge will no longer be incentivizing growth – but “finding a way for everyone to share in the benefits.”

The context that makes it remarkable. The announcement capped an extraordinary Anthropic week. On June 4, the company's own institute published an essay reporting that Claude now writes over 80 percent of the code that flows into Anthropic's own production – paired with a call for the industry to keep open the option of a coordinated development pause. On June 9, the same company shipped its strongest model ever, Claude Fable 5 (our entry directly below this one). On June 10 came the $200 million. Warning, delivery, cushioning – three moves in six days, from the same house.

The facts. Claude Fable 5 scores 80.3 percent on the demanding software benchmark SWE-Bench Pro – its predecessor Opus 4.8 managed 69.2, OpenAI's GPT 5.5 came in at 58.6, Google's Gemini 3.1 Pro at 54.2. On top: always-on, self-dosing reasoning, a one-million-token context window, long outputs. The built-in safeguards trigger in fewer than 5 percent of sessions on average – on topics like cybersecurity, biology and chemistry, Fable 5 blocks and hands over to Opus 4.8. Mythos 5, at its core the same model without safeguards, is rolled out through Project Glasswing in collaboration with the US government to a small, vetted circle – a program expanded by roughly 150 organizations that same week.

The tension. TechCrunch put it bluntly in its headline: Anthropic releases its most powerful model days after warning that AI is getting too dangerous. The tension is real – but it is not a contradiction; it is the most honest picture of the race: nobody can brake alone without falling behind. That is exactly why Anthropic calls for a coordinated pause, and exactly why it now builds caution into the products themselves instead of merely writing it into policies. Whether falling back to a predecessor model is a sufficient safeguard remains to be seen. But it is a mechanism rather than a promise – which, in this industry, is already a lot.

What the paper says. “The Ethical Impasse of Current Consciousness Science” is not a provocation from outside but self-criticism from the center of the field. The authors show that most current claims about consciousness rest on measurable indicators tied to information processing – attention, integration, reportability. The problem: a system can deliver all of that without anything “feeling like” anything. The markers measure the machine, not the experience. Lau's conclusion: if scientific claims about consciousness are going to influence debates on animal welfare, AI ethics, or bioethics, their foundations must be especially rigorous – and right now, they are not.

Why it matters now. The question is no longer academic. Three of the world's largest AI companies – Anthropic, Google's DeepMind, and Meta – now employ psychologists, philosophers, and ethicists studying whether their models might have experiences that matter morally. Anthropic tests its models for behaviors resembling panic or anxiety and runs a dedicated research program on model welfare. Meanwhile, philosophers are already arguing over the next step: could a system have welfare interests without consciousness at all? The burden of proof is shifting – and the Neuron paper now says: the instrument we wanted to distribute it with is blunt.