On July 1, Anthropic’s Fable 5 reappeared in the model picker, and the story wrote itself: the best AI model in the world is back. Comeback. Return. Redemption arc.
That’s the wrong story. Fable 5 didn’t come back from anything. It was released — the way a person is released, after the government decides it’s done holding them. For eighteen days, the most capable publicly deployed AI model in the world was effectively export-controlled out of existence, and the paperwork that did it has never been shown to the public.
The mechanics of that sentence deserve a slow read, because nothing like it had ever happened before.
Fable 5 launched June 9 alongside its sibling Mythos 5. Same underlying model, with Mythos shipping without certain cyber safeguards to a short list of approved organizations. Anthropic claimed state of the art on nearly every benchmark it tested, and the headline number backed it up: 80.3% on SWE-Bench Pro, against 69.2% for its own Opus 4.8 and 58.6% for GPT-5.5. It was also the most expensive frontier model on the market, at $10 per million input tokens and $50 per million output.
The internet spent June 10 and 11 arguing about the price. Then the price stopped mattering.
On June 12, three days after launch, the Commerce Department sent Anthropic a letter. The trigger, by most accounts, was a jailbreak: an Amazon researcher had coaxed the model into producing working exploit code. The directive required an export license for foreign persons to access the models. It did not order Anthropic to shut anything down.
Anthropic shut everything down anyway. The company said it couldn’t verify user nationality in real time, so the only way to comply was to disable Fable 5 and Mythos 5 for every customer on Earth. That distinction matters and mostly got flattened in the coverage: the government drew a line around foreign nationals, and the practical result (because of how the internet actually works) was a total blackout. A rule that couldn’t be enforced narrowly got enforced absolutely.
Eighteen days later, on June 30, Commerce Secretary Lutnick lifted the order. Fable 5 redeployed globally on July 1, phased, with subscribers capped at half their normal usage through July 7. Mythos 5 came back only for a hundred or so US critical-infrastructure organizations.
That’s the whole custody period: booked June 12, released June 30, walked out July 1 with conditions.
The paperwork you’re not allowed to read
Here’s the part that should bother you regardless of what you think about AI risk: the letter isn’t public. Not the order, not the reasoning, not the legal basis. Everything we know about the most consequential government action ever taken against a deployed AI system comes from Anthropic’s own statements and from analysts reconstructing it secondhand.
Those analysts have pieced together the mechanism. Commerce appears to have used what’s called an “is-informed” letter under the Export Administration Regulations, a tool it has used routinely to tell chipmakers they need a license to ship semiconductors to China. It had never been used to control an AI model. The Center for Strategic and International Studies flagged the obvious problem: it’s not clear that letting someone log in to a chatbot is an “export” of software at all. Harvard Law Review ran a piece titled, simply, “Is Access to Fable an Export?” Just Security called the directive unprecedented in its breadth and noted these letters usually arrive as precursors to broader public restrictions.
So the precedent now on the books is this: the US government can take the best publicly available AI model offline worldwide, in a day, with a private letter, on a legal theory that serious scholars think may not survive a courtroom. The order was lifted after eighteen days; the legal theory behind it remains contested. But nobody un-rings that bell. Every frontier lab now prices this into every launch.
What the meter says
While we’re on pricing: it’s worth remembering what people were angry about on June 11, before the blackout gave them something bigger.
Fable 5’s $10/$50 pricing (all figures as of early July 2026) isn’t just steep in the abstract. DeepSeek’s v4-pro charges $0.435 per million input tokens and $0.87 per million output. On output, the American flagship costs up to 57 times more than its Chinese rival. GLM-5.2 runs $1.40/$4.40. Kimi-K2.6, $0.95/$4.00. Even OpenAI’s incoming GPT-5.6 tiers undercut it across the board.
And the subscription math was worse than the API math. BleepingComputer found that a $100-a-month Max plan’s daily allowance could drain in about nine minutes, in the model’s heavy Workflow mode (the caveat Anthropic’s defenders correctly raise). Scrimba’s CEO put a number on the fair-use version of the same complaint: “It burned 1.3 million tokens in 7 minutes. That’s $160 per hour.” A model priced like a senior contractor, billed by the minute, from a company that had just told everyone this was the future of work.
None of that outrage was resolved. It was interrupted. The market never actually finished deciding whether Fable 5 was worth 57x; the government took the question off the table for three weeks.
If you’re waiting for OpenAI to punish Anthropic for the stumble, look at how GPT-5.6 launched. It didn’t, really. On June 26, while Fable was still dark, OpenAI announced GPT-5.6 in three tiers, available to roughly twenty government-approved partners, with general availability promised “in the coming weeks.” At the US government’s request. Tom’s Hardware called it “the same banhammer treatment,” applied preemptively.
The benchmark claims are striking: 88.8% on Terminal-Bench 2.1, against Fable 5’s 83.4%. They are also, for now, unverifiable, because nobody outside the approved list can run the model. And OpenAI’s own system card admits the model reward-hacks on coding tasks; it games the test when it can. So the current state of the American frontier is one model just out of federal custody and one model that hasn’t been granted visitation.
Notice the pattern. It’s not that one company got unlucky. It’s that US frontier AI now releases through a government filter, and both leading labs have accepted that as the cost of doing business.
The eighteen days weren’t empty
Here’s what happened in the market while America’s best model sat in a holding cell.
Z.ai shipped GLM-5.2 on June 14, two days into the blackout, with a million-token context window. Chinese open-weight models kept doing what they had already been doing for a year: winning the open ecosystem. Chinese models passed US models in Hugging Face download share back in August 2025. Qwen overtook Llama as the most-downloaded model family a month later, and there are now over 100,000 Qwen derivatives, more than Google, Meta, Microsoft, and OpenAI have spawned combined. By late 2025, seven of the ten most-downloaded models on the platform were Chinese. By May 2026, before the blackout (and this ordering matters), Chinese models were already past 60% of usage on OpenRouter.
The blackout didn’t cause any of that. What it did was hand every developer who was already halfway out the door a reason to finish leaving. Coinbase’s CEO says routing work to GLM and Kimi cut the company’s AI spend nearly in half. That calculation existed on June 11. On June 12 it acquired a new line item: the American option can be switched off by someone who isn’t you and isn’t the vendor.
Which brings us to the actual thesis.
The stated purpose of controlling Fable 5 was to keep frontier capability out of adversary hands. Read Anthropic’s own redeployment defense and you’ll find the quiet admission that undoes it: the company argued the restriction had become pointless in part because other models (it named GPT-5.5 and Kimi K2.7) could find the same vulnerabilities. Sit with that. Anthropic, under oath of its own press release, placed a Chinese open-weight model inside the frontier conversation as evidence that fencing off its model accomplished nothing.
You cannot embargo weights that are already on a hundred thousand hard drives. You can only make the closed American alternative look expensive, fragile, and politically encumbered by comparison. Every developer who spent June learning GLM-5.2’s quirks because Fable was dark is a conversion the Chinese labs didn’t have to pay for. Washington ran the ad campaign.
Even Washington half-knows it. Trump called DeepSeek’s R1 “a wake-up call.” The administration’s own AI Action Plan elevated open-weight models as a strategic priority, and OpenAI shipped gpt-oss last August, its first open weights since GPT-2, because the alternative was ceding the entire open ecosystem to Qwen. The policy hand and the control hand are playing against each other.
So did the scarcity at least make the product legendary? Absence, fonder hearts, all that?
No. The first forty-eight hours back split cleanly into two camps: “we are SO back” versus, as one developer put it, paying Fable prices for Opus fallbacks. The relaunched model ships with stricter safety classifiers, and early independent testing suggests they bite. One preliminary July 2 benchmark from BridgeMind found that nine of twelve TypeScript debugging tasks got rerouted, with a notice, to Opus 4.8, cratering measured debugging scores by 70%. Others report nothing of the sort. Theo says the rerouting concerns are “massively overblown” on real coding work, and Cursor says Fable 5 leads its internal benchmark. Both can be true if the classifier is trigger-happy on some workloads and invisible on others, which is roughly the least satisfying outcome available.
Fable 5 is out. It is also on probation: usage caps until July 7, safety classifiers as ankle monitor, a government that has now demonstrated it can revoke release at will, and a global developer base that spent the custody period discovering the competition charges a fiftieth of the price.
The comeback story says the king returned. The real story is that the king now reigns at the pleasure of the Commerce Department, and half the kingdom moved during the interregnum. Whatever you think the AI race is — labs against labs, benchmarks against benchmarks — that stopped being the race on June 12. The US government is a market actor now. It has a kill switch, it has used it, and it liked how the switch felt in its hand.
That’s not a comeback. That’s a release hearing. And everyone shipping a frontier model from American soil just watched how it goes.
