Three weeks ago I told you the government had switched off the most capable AI model in the country. This week it switched back on. That sounds like the story resolving itself. It didn't.
On Tuesday the Commerce Department lifted the export controls that forced Anthropic to pull Fable 5 and Mythos 5 offline for eighteen days. Fable 5 comes back to everyone today. But the terms of its return tell you the real story was never the outage. It was what Washington built while the lights were off, a standing review seat inside the release process for frontier models. And Anthropic is not the only company sitting through it. Four days before Fable came back, the same government quietly did the same thing to OpenAI, holding its newest model back from the public until a short list of approved partners could use it first.
Two labs, two flagship launches, the same Washington checkpoint inside the same ten days. Last month this looked like one company being punished over one jailbreak. This month it looks like a system, and neither company got a say in whether they wanted to be part of it.
Start with the timeline, because the dates do a lot of the explaining. Anthropic launched Fable 5 on June 9. Three days later, on Friday June 12 at 5.21pm Eastern, the Commerce Department sent a directive ordering the company to cut off any foreign national from Fable 5 and Mythos 5, inside the US or out, including Anthropic's own non-citizen employees. Because there was no way to sort users by nationality in real time, Anthropic pulled both models for everyone, worldwide. That is eighteen days offline. Access to Fable 5 begins restoring globally today. Mythos 5 got there first, reopened on June 26 to roughly 100 vetted companies and federal agencies under a letter from Commerce Secretary Howard Lutnick.
The trigger, revisited with a few more details than we had two weeks ago, was a jailbreak reported by researchers at Amazon, the same company whose CEO Andy Jassy reportedly first raised the alarm with the White House. The technique got Fable to read a codebase and identify, and in one case help exploit, a software vulnerability. Anthropic maintained throughout that the flaw was minor and already known, and that other public models including OpenAI's GPT-5.5 could do the same thing without triggering an export order. The government never published a technical justification. It didn't need to. The directive did the work on its own.
What actually ended the standoff is the part worth sitting with. Restoring Fable 5 didn't just mean flipping a switch back on. Anthropic is now deepening its cooperation with the government well beyond this one incident, agreeing to give designated agencies an early look at future frontier models and their safeguards before public release, to share threat intelligence, and to help build a common industry standard for scoring how dangerous a given jailbreak actually is, alongside Amazon, Microsoft, and Google. None of that is improvised. It flows directly out of the June 2 executive order on AI security, which asks companies to voluntarily hand the government up to 30 days of early access to covered frontier models before wider release. NSA, Treasury, and CISA have until August 1, sixty days after the order, to turn that into an actual classified benchmarking process.
Pressure built on both sides while the models sat dark. A coalition of information security leaders published an open letter pushing to get the controls lifted, arguing that AI oversight needs to be grounded in real testing and transparent rules rather than an unexplained directive. Critics inside and outside the industry also pointed out the obvious cost of the delay, that every day Fable and Mythos stayed offline was a day of runway handed to Chinese open source labs racing to close the gap. Anthropic, for its part, is doing all of this while confidentially preparing a public listing that could value the company near a trillion dollars, which means the government's willingness to switch its flagship product off, however briefly, was never just a technical dispute.
My take: Everyone is treating this as a relief story, and on the surface it is one. But look at what Anthropic actually agreed to in exchange for getting Fable back. Early government access to unreleased models. Shared threat intelligence. A jailbreak severity scale co-written with three of its biggest rivals. That is not a company getting its product back. That is a company that just helped design the machinery that will govern every future release, its own and everyone else's. Washington didn't blink first. It used eighteen days of standoff to install itself permanently inside the release pipeline, and Anthropic, still racing toward an IPO it cannot afford to jeopardize, had every reason to make that trade look like a partnership rather than a surrender.
On the same Friday that Mythos 5 got its partial reprieve, OpenAI released its newest model family, GPT-5.6, in three tiers it calls Sol, Terra, and Luna. Sol is the flagship. None of the three launched to the public. OpenAI limited the rollout to roughly 20 partner organizations whose access had been cleared in advance by the government, with no ChatGPT toggle and no public waitlist. Axios called it the first time Washington has asked an American AI company to restrict a model's release before it ever shipped, rather than recall one after the fact.
The backstory is close to a mirror of Anthropic's. The White House's Office of the National Cyber Director and Office of Science and Technology Policy asked OpenAI to hold GPT-5.6 back. Commerce Secretary Lutnick reportedly pushed for sign-off from additional agencies before even the limited preview could go ahead, and Sam Altman discussed the model with Lutnick directly before Friday's launch. OpenAI complied, but made its unhappiness plain in its own announcement. "We don't believe this kind of government access process should become the long-term default," the company wrote, framing the restriction as a short-term step toward a broader release it expects within weeks.
The model itself is a serious release. OpenAI says Sol edges out Anthropic's Mythos 5 on its own coding benchmark while using roughly a third of the output tokens, and it's priced at $5 per million input tokens and $30 per million output, half that for Terra and a fraction for Luna. But the pricing and the benchmarks are not what will be remembered about this launch. What will be remembered is that two competing labs, in the same ten day window, both routed their most capable models through the same Washington office before a single outside user got to try them.
Neither company was legally required to do this. The June 2 executive order is explicitly framed as voluntary, and creates no formal licensing requirement on paper. Dean Ball, a former White House AI adviser now joining OpenAI, told TechCrunch the arrangement already functions as "a de facto involuntary licensing regime for frontier AI," warning that without published safety standards, review timelines could stretch out indefinitely and hand real advantage to competitors overseas. Staff in a short list of cleared countries, including the UK and Australia, can use GPT-5.6 even where their employer is US based. Everyone else waits.
My take: This is the detail that turns last issue's story from an Anthropic problem into an industry structure. If only Fable 5 had been pulled, you could file it under bad luck, a narrow jailbreak that spooked the wrong official on the wrong Friday. But GPT-5.6 was never accused of anything. OpenAI submitted to the same checkpoint pre-emptively, because that's apparently what happens now once a model crosses some unpublished capability line. Nobody outside a handful of federal offices knows exactly where that line sits, which is precisely the problem. A voluntary framework that two rival labs both treat as mandatory isn't voluntary. It's just informal, and informal is worse, because there's no appeal, no published standard, and no clock you can count down except the one Washington sets for itself.
Two issues ago I wrote that the labs building the frontier don't actually control it. This week adds a wrinkle. The gates are multiplying, and they're stacking on top of each other rather than replacing one another. Washington now sits inside the release calendar for the most capable models, whether a lab wants it there or not. Underneath that, the physical layer, the actual buildings and chips, keeps consolidating into fewer hands too, Digital Realty buying out its own partner, SpaceX turning a training cluster into a landlord with four paying tenants. A model you can access still has to run somewhere, and increasingly that somewhere is owned by the same small list of names.
Then there's a third gate nobody was tracking closely until this week, use itself. Anthropic's case against Alibaba is really an argument that querying a competitor's model too much, too fast, in too coordinated a way, is its own kind of violation, not of a law exactly, but of a norm serious enough to bring to the Senate. Put that next to the government's pre-release reviews and you get a strange picture. You can be gated for building something too capable. You can be gated for using something too aggressively. The space in between, where you build responsibly and use moderately and still somehow avoid every checkpoint, keeps getting narrower.
None of this makes the individual decisions wrong. A government wanting to see a dangerous model before the public does isn't unreasonable on its face, and a company wanting to protect years of training investment from a scraping campaign isn't either. But stack all three gates, government clearance, compute ownership, and now usage scrutiny, and you get an industry where fewer and fewer paths run all the way from a good idea to a shipped product without someone else's approval sitting somewhere in the middle. That's the story underneath both of this week's headlines, and it's the one that will still be running long after Fable 5 and GPT-5.6 are both back to normal.
The Ratepayer Protection Act has now been parked in the Senate Rules and Operations Committee since June 8, with no floor action since. It passed the House 69 to 44 back on June 3, nearly a month ago, and the wait since has been long enough that it's worth asking whether this is deliberate stalling or ordinary summer scheduling. Either way, the bill that would make data centers pay for their own grid upgrades is still sitting exactly where it was two issues ago.
While SB 730 waits, the state found another venue to move in. The Energy Policy Council's Load Growth Task Force met last Thursday to weigh new rules for the state's largest electricity users, including minimum monthly bills, exit fees if a promised data center never materializes, and "bring your own capacity" programs that would let a large customer build or contract for its own power instead of leaning on Duke's grid. It's a smaller, quieter echo of the FERC order I described two issues ago, which told the nation's big regional grid operators to rewrite how they handle massive new loads. The difference is that FERC's order barely touches North Carolina, since Duke runs its own grid outside any regional transmission organization. Whatever protection gets built here has to come out of Raleigh, not Washington, which is exactly why a stalled Senate bill matters more here than it might in a state that's covered by a federal fast track.
Duke's 18% rate request is still working through the Utilities Commission, and Attorney General Jeff Jackson's counter-proposal, which he says would save North Carolinians $1.4 billion over two years, is still on the table alongside it. None of that has moved much in the last two weeks either. The state's slowest-moving pieces and its fastest-moving ones are sitting right next to each other right now, a Senate bill gathering dust and a federal AI checkpoint reshaping model access twice in ten days.
For Triangle builders, the lesson from this issue's main stories lands close to home. It used to be enough to worry about which model your product depended on. Now you're also depending on which government office cleared that model for release, and on which data center company owns the building it runs in, and increasingly on how heavily your own usage patterns might get read by that vendor's legal team. None of those dependencies are things a small team here can control. What you can control is not betting the whole company on any single one of them holding steady, because this month alone, every layer moved.
That's the week the frontier reopened on Washington's terms. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. If your team got Fable 5 or Mythos 5 back this week, or if you're one of the twenty or so shops sitting on GPT-5.6 access while everyone else waits, tell me what the gated experience has actually been like day to day. I'm collecting real accounts of what it's like to build on a model that's stuck behind a government checkpoint, for a future issue.
P.P.S. Forward this to the person who thought the Fable shutdown was a one-time story. This week a second company walked straight into the same checkpoint on purpose.