The switch exists now

The controls being lifted is not the story. The controls having been possible is. For two weeks in June, the US Department of Commerce demonstrated it can toggle a commercially deployed frontier model off the global market by letter, and every product roadmap in the industry now has to price that in.

What actually happened. On 1 July, Commerce Secretary Howard Lutnick announced the lifting of export controls on Anthropic's Fable 5 and Mythos 5, imposed roughly two weeks earlier under an "Is Informed" letter from the Bureau of Industry and Security. Fable 5 returned globally across Claude.ai, Claude Code and Claude Cowork within a day. Mythos 5 did not: it remains accessible only to vetted US organisations through Anthropic's Glasswing programme, with case-by-case expansion. Anthropic offered affected users up to 50% weekly usage credit through 7 July as commercial cushioning.¹²³

The read most coverage settled on. A short, reversible episode. Commerce acted, industry complied, negotiation followed, controls lifted, life resumed. On the surface this is the correct summary and I am not going to pretend otherwise. What I want to argue is that the surface is misleading about two durable things: the option Commerce has just proved it holds, and the access architecture Anthropic has quietly locked in during the fortnight everyone was watching the wrong door.

The option is now real

Before mid-June, export controls on frontier AI models were a theoretical instrument. The Export Control Reform Act clearly permitted it; nobody had actually done it to a deployed product with paying international customers. The "Is Informed" letter changed that. Commerce can now suspend access to a specific model, by name, with a mechanism that took effect immediately and stayed in effect for two weeks. The reversal came with commitments from Anthropic on enhanced risk monitoring, information sharing, and coordination on future releases.⁴

The precedent runs in both directions. Labs now know Commerce will pull the lever when it thinks the cybersecurity profile of a model warrants it. Commerce now knows the political and commercial cost of pulling the lever is survivable inside a two-week window. Both parties learned that a negotiated re-authorisation, with monitoring commitments and a vetted-access carve-out, is the resolution path. That is not an episode. That is a template.

The credit is the tell. A company offering half of a week's usage back to its customers is a company that has just been reminded its product has a regulatory dependency it does not control. Every future model launch — from Anthropic, but also from OpenAI, Google DeepMind, and anyone else training at frontier scale — now factors in a non-zero probability of temporary suspension between announcement and steady-state availability. That probability will be priced into launch communications, customer contracts, and, in due course, insurance products that do not yet exist.

Glasswing is the structural shift

Fable 5 came back for everyone. Mythos 5 did not. It is available through Glasswing, a vetted-access programme for US organisations, expanding case by case. In the pre-June world, capability tiering happened by geography (export rules), by price (enterprise vs consumer), and by product surface (API vs chat). Glasswing adds a fourth axis: organisational vetting status. A US hospital system that passes Glasswing review gets a materially more capable model than a US hospital system that does not, and than any organisation outside the United States regardless of size or intent.

This is what "sovereign AI" actually looks like when it stops being a phrase in a keynote and becomes an access-control list. The Constellation write-up is careful to note the case-by-case expansion, and the honest reading is that Glasswing may widen substantially over the next year.³ But it may also not. Either way, the mechanism now exists: a two-tier regime where the top tier is gated by a review process the vendor administers on terms it has agreed with Commerce. That is a durable change, and it happened while the headlines were focused on whether Fable 5 would come back.

The joint framework and what it is really for

Anthropic, Amazon, Microsoft, and Google have jointly developed a jailbreak-severity framework, positioned as giving Commerce a structured technical basis for future decisions.⁴ I want to name the tension without pretending to resolve it. On one reading, this is exactly what a mature industry should do: propose objective criteria so that regulatory action follows a rubric rather than a phone call. On another reading, it is the four organisations with the largest commercial stake in where the lines fall co-authoring the ruler that will measure them.

Both readings are true simultaneously. The framework will probably improve on ad hoc discretion, which is a low bar the June episode cleared with room to spare. It will also entrench the position of the incumbents who wrote it, because any smaller lab or open-weights project will have to argue against a scoring system it did not help design. The critics who pointed out during the blackout that two weeks of Western-frontier unavailability is a gift to Chinese open-source competitors were making a real point.²⁵ The framework is, among other things, the industry's answer to that argument: a way to make "you are helping the competition" legible to Commerce as a number rather than a lobbying line.

What to watch. Not whether more controls are imposed. Whether Glasswing's vetted tier expands or narrows over the next two quarters, and whether any non-Anthropic lab adopts a comparable structure. If Mythos-tier capability becomes routinely available only through vendor-administered vetting programmes across the frontier, the industry has quietly moved from a single global market for AI capability to a segmented one, with the segmentation controlled by the vendors under regulatory supervision. That is a bigger change than a two-week outage, and it is the change this episode actually delivered.

The switch exists now. The tier exists now. The rulebook is being written by the four companies with the most to lose from it being written badly, and the most to gain from it being written at all. None of that is undone by Fable 5 being back online this morning.

Footnotes

¶