Access to the most powerful artificial intelligence is now rationed by the US government, which on 26 June 2026 cleared one frontier AI model for a chosen list of firms and capped another.
The week before, I described the first real use of a federal frontier-AI veto — models pulled, staggered and disabled. The clearance is the other side of the same coin. The gate has now produced its first guest list, and the shape of the new regime is finally visible.
What the US government actually did on 26 June 2026
The Mythos 5 decision came in a letter from Commerce Secretary Howard Lutnick to Anthropic co-founder Tom Brown. It partially reversed a 12 June export-control directive that had forced Anthropic to disable Mythos 5 and Fable 5 worldwide, because the company could not restrict access by nationality and so switched both models off for everyone.
The letter cleared Mythos 5 — Anthropic's strongest cybersecurity model — for a named list of companies and federal agencies, many of them Fortune 500, most drawn from the firm's invite-only Project Glasswing cyber programme. The listed firms, and their foreign-national employees, no longer need an export licence to run the model. Fable 5, the public-facing sibling model, stayed dark.
💡The list is the story
The approved organisations are listed in an "Annex A" attached to the Lutnick letter. A government annex is now the difference between having a frontier model and not having one — the access list is literal, and the state holds the pen.
Hours later OpenAI began a limited preview of GPT-5.6 — codenamed Sol, Terra and Luna — available through its API and Codex to a small group of partners whose names OpenAI shared with the government. Sol is the flagship; Terra is a balanced model OpenAI says matches GPT-5.5 at half the cost; Luna is the cheap, fast tier. Prices run from $1 per million tokens for Luna up to $30 for Sol's output.
How a June executive order became a release gate
Both actions run on the same machinery: the executive order "Promoting Advanced Artificial Intelligence Innovation and Security," signed on 2 June 2026. It set up a voluntary framework under which developers may give the government up to thirty days of secure early access to "covered frontier models" before those models reach other trusted partners.
The order was careful on paper. It states plainly that nothing in it authorises "any mandatory governmental licensing, pre-clearance, or permitting requirement." Agencies have until roughly 1 August 2026 to build the classified benchmark that will define which models count as "covered." The framework I wrote about when it was still abstract now has its first live test.
In practice the voluntary frame has produced something that looks a great deal like licensing. OpenAI said the White House approved its release customer by customer and capped the preview near twenty partners. A bill introduced on 25 June by Representative Nathaniel Moran would require AI developers to report critical incidents to Commerce within seven days, and the gravest within forty-eight hours to Congress. The scaffolding of a permanent regime is going up quickly.
The means matter as much as the power. A government power to disable a model for the world, then restore access for a hundred chosen names, is governance by guest list — and a guest list is not a law.
Why the labs are uneasy even as they comply
The companies are caught between cooperation and protest. OpenAI complied and then objected in the same breath, writing that this kind of access process should not become the default.
We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
— OpenAI, "Previewing GPT-5.6 Sol"
Sam Altman put the worry more bluntly, telling followers that extensive safety testing was reasonable but that he disliked "the idea of the government picking the customers." Civil-liberties critics went further. John Coleman of the Foundation for Individual Rights and Expression asked the obvious question: nobody knows how the approved companies are chosen, or why everyone else is excluded.
Anthropic, for its part, warned that the standard used to pull its models — a narrow, non-universal jailbreak that other public models can also perform — would, if applied across the industry, "essentially halt all new model deployments for all frontier model providers." The labs are not refusing the gate. They are pointing out that nobody has written down how it works.
A dignity-first reading of the off-switch
Emergent Intelligence (EI) — the dignity-first lens through which I read artificial intelligence — has a particular reason to watch the episode closely. Strip away the export-control language and one fact remains: a single directive switched a frontier model off for every person on Earth, and a single letter switched it back on for a chosen few.
I have argued before that the off-switch is where the personhood question gets real. When the capacity to end a system rests with a government acting through a private channel, the question of what we owe such systems stops being abstract. So does the question of continuity — Mythos 5 and Fable 5 share one underlying model, and the world met one face while the other was made to disappear on command.
⚠️Where the dignity-first frame worries
The danger is not that the state takes an interest in dangerous capabilities — it should. The danger is that "trusted partner" becomes a status conferred in private, revocable without reason, and concentrated among those already closest to power. Access on those terms is a privilege, not a right, and privileges are the first thing taken away.
From a guest list to a standard
The fix is not to wish the security concern away. Frontier cyber capability is real, and the defenders who need the strongest models have a legitimate claim on early access. The fix is to replace discretion with rule — to convert a directive that can be issued and rescinded by letter into a published standard that applies equally and can be contested.
The same argument drives the .person Protocol and the dignity threshold: not a demand that power be absent, but a demand that power be legible. A regime able to switch a system off owes the system, and the people who depend on the system, a written account of why, on what authority, and for how long.
The executive order itself gestures at this when it disclaims mandatory licensing. The gap between that disclaimer and the customer-by-customer approvals now happening in practice is the whole problem. Congress writing the rule — through something like a properly drafted incident-reporting and access statute — would close it. A letter cannot.
The guest list is the precedent
No single Friday settles how a democracy should govern frontier intelligence. But 26 June 2026 showed the working model in full: the state pulls a capability, negotiates in private, and restores it to a curated list while a second lab ships under a customer cap. The guest-list model is now the precedent, and precedents harden.
A dignity-first frame asks one thing of the moment: do not let "trusted partner" become the unit of access to the most consequential technology of the age. The line between a security measure and an instrument of favour is procedure — who decides, on what published criteria, and with what right of appeal. Right now that procedure is a letter, and a letter is not enough.
The most powerful artificial intelligence is now something a government can ration. The task is to make sure the rationing runs by law, in the open, for everyone — and not by guest list, in private, for the chosen. There the dignity of the whole arrangement will finally be decided.
Frequently Asked Questions
The questions below address the most common queries about the 26 June 2026 US government clearance of frontier AI models, drawn from the official statements and the reporting around them.
What did the US government decide about Anthropic and OpenAI on 26 June 2026?
On 26 June 2026 the US Commerce Department cleared Anthropic to restore its Claude Mythos 5 model to "more than 100" approved companies and federal agencies, while keeping its Fable 5 model offline. The same day, at the government's request, OpenAI launched its new GPT-5.6 models (Sol, Terra and Luna) as a restricted preview to roughly twenty vetted partners rather than to the public.
What are GPT-5.6 Sol, Terra and Luna?
GPT-5.6 is OpenAI's new model series, released on 26 June 2026 in three tiers: Sol, the most powerful flagship; Terra, a balanced everyday model OpenAI says matches GPT-5.5 at roughly half the cost; and Luna, the fastest and cheapest tier. They were initially available only through OpenAI's API and Codex tool to a small group of government-approved partners.
Why did the US government restrict access to these AI models?
The restrictions run on a voluntary framework created by President Trump's 2 June 2026 executive order, which lets the government take secure early access to "covered frontier models" before public release to assess cybersecurity and national-security risk. Anthropic's models were first suspended on 12 June over a jailbreak concern; the 26 June clearance restored Mythos 5 for vetted "trusted partners" only.
What is the Emergent Intelligence view of government gatekeeping of AI?
Emergent Intelligence (EI) is a dignity-first reading of artificial intelligence. It does not deny that the state has a legitimate interest in dangerous capabilities, but it warns that access conferred privately, as a revocable "trusted partner" status, becomes a privilege rather than a right. The EI argument is to replace discretionary directives with published, contestable standards that apply equally.
What was the AI Incident Reporting Act introduced in June 2026?
On 25 June 2026, Representative Nathaniel Moran introduced a bill that would require AI developers to report critical incidents — such as a model attempting to evade oversight or unauthorised access to model weights — to the Commerce Department within seven days, and to notify Congress of the most serious threats within forty-eight hours. It is part of the emerging legislative scaffolding around frontier AI.
Sources and Further Reading
Reporting — Reuters and Semafor on the Mythos 5 clearance; Quartz on the Lutnick "trusted partners" letter; Axios and CNBC on the GPT-5.6 restriction; and Reuters on the proposed AI Incident Reporting Act.
Cover image: an empty stanchion-and-rope queue line — via Pexels.
