Model access is now a government-approval workflow, not just a product toggle. The proof arrived in three dated steps. [1]

On June 12, the Commerce Department applied an export-control order to Anthropic's Fable 5 and Mythos 5 models, restricting access by foreign nationals — and because the company had no reliable way to verify nationality in real time, it suspended both models for everyone on earth. [2] On June 26, Washington approved restoring Mythos 5 for a set of US organizations. On June 30, Commerce Secretary Howard Lutnick announced the controls were lifted entirely; Anthropic restored Fable 5 globally on Wednesday, with usage caps through July 7 and fresh classifiers trained against the jailbreak technique that started the episode. [2] The company's own announcement of restored access confirmed Commerce approval as the trigger. [3]

Nineteen days, three decisions, one product line dark for most of it. That is the receipt, and it reorders who runs commercial AI: a model release now ships with an implicit desk officer. For Anthropic's customers the episode was not an abstraction — enterprise builds that assumed frontier access spent two weeks rewriting prompts around older models, and every roadmap that assumed uninterrupted API availability now carries a clause nobody drafted. [1] [2]

The substance underneath matters as much as the sequence. The original order followed a threat-intelligence report describing a method to bypass Fable's safeguards and expose offensive cyber capabilities; Commerce treated model access as something it could license. Legal scholars spent the interim noting that federal advisory opinions since 2009 said remote access was not an export at all, which suggests the theory will be tested — one customer with Canadian employees has already sued. [2]

The resolution's fine print shows what compliance now costs. Fable 5 returns globally but capped at half of weekly usage limits through July 7 for most paid plans, after which it moves to usage credits; cloud partners get re-enabled as fast as coordination allows; Mythos 5 stays restricted to an approved set of US critical-infrastructure organizations pending expansion of the government-coordinated Glasswing program. Anthropic paired restoration with newly trained classifiers targeting the reported jailbreak and proposed a shared industry framework for judging jailbreak severity — a standard that did not exist when a single threat report froze two frontier models worldwide. [2]

Here is the gap. X frames the episode as either censorship of a leading lab or overdue discipline of one; MSM reports each announcement without connecting them into precedent. The verifiable middle: the dates above, plus what happens next time. OpenAI's parallel preview of its next model to government first, releasing only to "trusted partners," shows the workflow spreading beyond one company. [2] When two competitors treat pre-clearance as normal, the market has adopted the rule regardless of what any court later says.

The paper's position: watch the approval latency, not the press releases. Nineteen days for this round; if the next frontier release clears faster, the loop is institutionalizing quietly, and commerce in models will run at the speed of its slowest regulator. Every enterprise contract signed this month should be read with that dependency priced in.

-- THEO KAPLAN, San Francisco