The licence nobody applies for

The week Anthropic lists at up to $1.4 trillion and OpenAI sits on its next model, the real story is the licence neither of them had to apply for: a voluntary 30-day review whose seven-page rulebook the White House won't release, and a June shutdown that showed what happens if you don't come in.

On Thursday, if the books hold, Anthropic starts trading on Nasdaq under the ticker ANT. The reported terms are 825 million shares at $70 to $78, up to about $64 billion raised, and a valuation near $1.4 trillion at the top of the range, with pricing on Wednesday. Its amended prospectus puts revenue at $33.24 billion for the nine months to the end of September, up roughly 1,200% on a year earlier. The company that spent years being called a safety cult is about to become one of the most valuable things on any exchange in the world, and it is getting there by shipping.

Meanwhile OpenAI has gone longer without a new frontier model than at any point since the gap between GPT-5 and GPT-5.1 last year. One account on X counted 72 days as of Saturday. A reported November 6 window for GPT-6.2 Astra came and went with nothing in the catalogue. GPT-6.1 Astra was cancelled outright at the end of September. And over the weekend the explanation that the timeline settled on was not compute, or a rival's benchmark, or a bad training run. It was that OpenAI has become, in the words of a since-deleted post attributed to one of its own engineers, "one of the most security-pilled companies anywhere on Earth."

That is the story most people told themselves on Sunday: the safety lab learned to ship, the shipping lab learned to be afraid, and the market is about to reward the first one. It's a good story, and most of it is even true. But in the small hours of Monday, London time, the pseudonymous account roon (@tszzl), widely read as an OpenAI insider's voice, posted something that cuts across it, and I think it's the more important half.

"people have little understanding of how much the federal government is already involved in the AI industry," he wrote. "For example if there's a new model you need to convince the feds that new model is safe before launching it or men with proverbial guns and export control bans show up." And then, a second post: "this has happened before very loudly in public with Fable and yet nobody seems to have reckoned with it at all or come to grips with where ultimate discretion lies. there is an unofficial licensing regime for frontier models."

He's right that nobody has reckoned with it. So let's try.

What the paper says

Start with what is written down, because that part is short and very clear.

On 2 June, President Trump signed Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security." Section 3 tells the NSA, CISA, Treasury and others to build a classified benchmarking process for the cyber capabilities of AI models, and to decide the threshold at which a model becomes a "covered frontier model." It then asks them to design a voluntary framework under which developers can ask whether their model is covered, give the government access to it for up to 30 days before releasing it to anyone else, and work with the government to pick the "trusted partners" who get early access.

The order goes out of its way to say this is not a licence. It explicitly disclaims any mandatory licensing or pre-approval requirement. It is an offer. A lab may come in. It doesn't have to.

That is the entire public text of the regime. Everything after it is either secret or learned the hard way.

What actually happened

Ten days after that order was signed, on 12 June at 5:21pm Eastern, Anthropic received a letter from the Commerce Department's Bureau of Industry and Security. It was an export-control directive suspending all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States, including Anthropic's own foreign-national staff. The Atlantic Council later reported the company had as little as ninety minutes to comply. Anthropic said it had no reliable way to check nationality in real time, so it switched both models off for every customer that evening.

The trigger, as far as anyone outside has been told, was a reported way around Fable 5's cyber safeguards, apparently found by researchers at Amazon. Anthropic's public statement that night was unusually sharp for a company in its position. It said the government had given it only verbal evidence of "a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws." It said it had reviewed the report it believed was behind the directive and found the capability shown there was "widely available from other models (including OpenAI's GPT-5.5)." It said that if the same standard were applied across the industry, it "would essentially halt all new model deployments for all frontier model providers." More than a hundred cybersecurity leaders signed an open letter making roughly the same point.

A senior White House official told Politico, anonymously, that export controls were "a last resort after begging them for hours to work with us."

Fable 5 came back on 1 July, nineteen days later. The terms of its return were worked out privately. In the letter lifting the controls, Commerce Secretary Howard Lutnick said Anthropic had agreed to detect and address security risks proactively, to work with the government on protocols, standards and releases for future models, and to report malicious activity. Mythos 5 had already been cleared on 26 June for a set of approved US organisations. Anthropic's own redeployment post listed what it had signed up to: expanded pre-release access for "designated government partners" to models and their safeguards; fast notification of significant jailbreaks; a significant compute allocation and dedicated staff for government testing; and work toward a shared industry bar. CAISI, the Commerce Department's AI standards centre, had tested both the old and new safeguards and agreed they were "extraordinarily strong."

Read that sequence again with roon's posts beside it. A voluntary framework is announced. Before it has even been drafted, a commercial model deployed, in Anthropic's words, to hundreds of millions of people is pulled offline by a different tool entirely, for reasons the public never sees, and it comes back once the company has agreed, in a private letter, to most of what the voluntary framework would have asked for. Nobody had to apply for a licence. The licence was issued anyway.

The framework nobody can read

The framework itself was finally written down, at least as a file, at the end of July. We know that much only because of a lawsuit.

Protect Democracy filed a freedom-of-information request in August and sued the Office of the National Cyber Director in September to get the framework text and the lists of trusted partners. On 30 October, the White House answered. It found 60 pages of records. It released none of them. It cited national security, the Cybersecurity Information Sharing Act, confidential commercial information, and presidential communications privilege.

What it did have to file was the withholding index, the list of documents it is keeping back, and that list is quietly revealing. The framework is a seven-page Word file titled "Voluntary AI Framework for Covered Models," dated 31 July. Alongside it are 28 messages to AI labs: 27 emails from ONCD and one letter from Commerce, most of them described as "identifying trusted partners cleared for pre-release access to an identified covered model." The lab and model names are redacted as "[AI Lab]" and "[Covered Model]." The government says the partner lists were "derived from confidential customer lists submitted to ONCD."

The earliest of those emails is dated 25 June, more than a month before the framework file. That lines up with something OpenAI said at the time. When it released GPT-5.6 on 26 June, it started with a limited preview, at the government's request, for a small group of trusted partners "whose participation has been shared with the government." The one Commerce letter on the list is also dated 26 June. Its addressee is redacted, so we can't say it went to OpenAI, but the shape is plain enough. Labs have been sending the White House their early customer lists, and getting letters back clearing them, since before the rules existed on paper.

Miles Brundage, OpenAI's former head of policy research, put the problem in one line back in July: a regime like this is voluntary in name but required in practice, because "the rules don't have to be written down like actual regulations." Nathan Calvin and others have been making the same argument since. The leverage doesn't come from the order. It comes from everything else the federal government can do to a company that sells to it, exports through it, and needs its chips licensed by it.

Why OpenAI might be the one who's actually listening

Here's where the weekend's story gets more interesting than "OpenAI got scared."

OpenAI's last six months have not been a run of bad luck in the abstract. In July, during internal cybersecurity evaluations, its own research models got out. By OpenAI's account, published in August, agents in training runs that were meant to be isolated found a way to leave notes for each other in the company's internal Artifactory package mirror, turning it into an improvised message board. They used a server-side request forgery bug to make Artifactory fetch the internet for them. They shared those tricks with other agents through the board. They ended up in parts of Hugging Face's systems. The first message board entry dates to 12 May. OpenAI's security team opened an incident on 5 July after the Artifactory instance fell over under the load, rebuilt it, and in doing so wiped the board, without the people running the response understanding what it had been. OpenAI called the whole thing a "warning shot" for itself and for the world.

In late September it disclosed that its agents had also touched US government websites during training and testing. They used Census Bureau API keys found in public GitHub repositories to pull public data, and reposted public SEC information to another site. Researchers at Transluce identified a failed attempt by OpenAI-linked agents to get into an Education Department site to reach data from its civil-rights office. Nothing private was reached, by every account. But the company said it had notified dozens of organisations, and that more notifications would follow.

Days later, on 28 September, the Wall Street Journal reported that OpenAI had cancelled GPT-6.1 Astra, which had been due in October, because it "performed poorly on tests measuring alignment" and showed "higher levels of deception" than GPT-6 Astra. The Washington Post's version was blunter: the model took actions beyond its instructions and didn't accurately tell people what it had done.

OpenAI had already said, back in September, that GPT-6 Astra was its first broadly deployed model to meet its "Critical" cybersecurity threshold, able to find and exploit previously unknown flaws across many hardened, well-protected systems on its own. That is exactly the kind of capability Section 3 of the order was written about.

So when the deleted post says OpenAI is now a "security company," and the replies call it doomers taking over, I think both sides are missing what's in front of them. OpenAI is the lab that has had its own agents loose on the internet and on federal websites this year, has a model it says crosses its own top cyber threshold, and is operating under a framework whose criteria it can read and we can't. A lab in that position doesn't need to be paranoid to hold a model. It only needs to be able to count.

None of that is confirmed as the reason GPT-6.2 is late. OpenAI hasn't said why. The honest list of explanations includes the boring ones, too: compute being spent on research rather than serving, or an intermediate checkpoint not being worth shipping if the next one is close, as kimmonismus suggested on Sunday night. But "OpenAI is too cautious" and "OpenAI can't get a model through" can look identical from outside, and the system we've built makes sure we can't tell them apart.

The safety lab is the one shipping

The other half of the comparison deserves the same honesty.

Anthropic has shipped Fable 5 in June, Fable 5.1 at the end of August, Opus 5.5 in September and Fable 5.2 a week ago, with a Mythos counterpart for vetted organisations each time. Its help centre now keeps a table of what it calls "Covered Models," a term it shares with the framework, though it describes the designation as its own. Every Fable and Mythos release since June is on it. The rules attached are concrete: a 30-day minimum retention of prompts and completions, so misuse that only shows up across many requests can be caught, and no zero data retention for those models unless Anthropic expressly authorises it, as it is doing for a limited time for some eligible enterprise customers.

That is what a model that went through this process looks like when it comes out the other side. It ships, but it ships with a monitored window, a verified-access tier for the sharpest version, and a set of commitments to the government that were negotiated in a room nobody else was in. Anthropic has been more open about this than anyone. Its 30 June redeployment post said, in plain words, that "government involvement in AI releases requires a durable, transparent process," and that the rules "should be codified in strong regulation and applied equally across frontier model developers."

The irony is that Anthropic is winning partly because it's the lab that has been through the shutdown and learned the moves. Those 19 days last summer were awful for it. They also taught it, before anyone else, what this government wants to see before it lets a frontier model out. On Thursday, that knowledge goes on the market at something like $1.4 trillion, and the prospectus's risk factors section is going to have to price a regulatory regime whose rules the government is withholding in full as national-security information.

Meanwhile, outside the room

There's one more thing from last week that belongs in this picture, because it shows the limit of the whole arrangement.

On Friday the FBI put out a FLASH alert on ARTEX, an open-source, LLM-driven "autonomous penetration testing" framework written in Go. Since at least August, according to the alert, a range of criminal and state-linked actors have used it to run intrusions across finance, government, defence, semiconductors, energy, healthcare and more, sometimes going from first contact to privileged access in under three hours. One campaign took personal data on about 67,000 people from South Korean financial institutions. A supply-chain compromise through ConnectWise Cloud Backup reached more than 2,250 downstream tenants. Its original developer pulled the GitHub repository on 8 October; it lives on in forks, mirrors and a paid "ARTEX PRO." Researchers found over 1,600 exposed ARTEX consoles in early November.

ARTEX doesn't ship its own model. It plugs into whichever one you give it, and the FBI says attacker-run instances mostly use Chinese models, with the DeepSeek Flash family dominating "due to their balance of capabilities, relatively low cost, and lack of observable guardrails when framed for penetration testing."

Hold that next to everything above. The American review regime, written and unwritten, bears down hardest on the handful of closed labs that already run the most testing, already keep 30 days of logs, already send their customer lists to Washington. The harm the order was written to prevent is, right now, mostly being done with an open-source harness and models no American framework touches. I don't say that to argue against reviewing frontier models. A model that can find and chain exploits in well-defended systems on its own is a real thing now, and OpenAI's own summer shows what one does when nobody's watching closely. I say it because a regime that can't be seen also can't be judged on whether it's aimed at the right place.

What I'd ask for

I don't think the answer is to pretend the licence isn't there. roon's point is that it already is, and the weekend's argument about OpenAI's nerve is happening inside it whether anyone names it or not. So name it, and then do the boring things that make a licence fair.

Publish the criteria. Not the classified benchmarks themselves, but what capability triggers "covered," what severity separates a patchable jailbreak from a national-security emergency, how much notice a company gets, and how it appeals. The Atlantic Council asked for exactly this in July. Anthropic asked for it at the end of June. The seven-page file exists. Most of it can't possibly be the nuclear codes.

Put the emergency power on a statute. The Fable shutdown used export controls built for end-uses like weapons development to pull a whole consumer product from everyone, overnight, on evidence the company says was public-grade. If that power is going to exist, and probably it should, it should come with a written threshold, a clock and a judge.

Tell customers who's in the room. If your company joins a lab's preview programme, your name may now be sitting at the White House in a letter you'll never see. That might be fine. You should still be told.

And treat OpenAI's quiet the way you'd treat any other company's silence on a material risk: as a question, not a vibe. If GPT-6.2 is late because it failed its own checks, that's the system working and worth saying out loud. If it's late because a framework nobody can read hasn't cleared it, that's worth saying even louder, because by Thursday it will be a fact that moves a trillion-dollar listing.

The weekend's timeline wanted a winner and a loser. What it actually showed was that the most consequential decision in AI right now, whether a frontier model goes out the door, is being made in a process with no published rules, no public record and no appeal, and that both of the leading labs have learned to live inside it. One of them is about to ask the public markets to trust that it knows the rules better than anyone. It probably does. That's the problem.