In late June 2026, Washington showed it could gate the release of frontier AI models, holding back OpenAI’s GPT-5.6 Sol before launch and forcing Anthropic to take Claude Fable 5 and Mythos 5 offline worldwide. The two labs answered through different instruments: OpenAI offered Washington a 5% equity stake, while Anthropic complied with an export-control directive that pulled its models offline. The wider story of government-gated AI releases turns on a single question: what does a lab owe a government that can gate its release?
Why did the US government ask OpenAI to restrict the GPT-5.6 Sol release?
The US government asked OpenAI to hold GPT-5.6 Sol because its cybersecurity capability crossed the “High” threshold under OpenAI’s Preparedness Framework. Commerce Secretary Howard Lutnick warned Sam Altman against releasing it without approval. It was a pre-release hold: Sol launched on 26 June to roughly 20 trusted organisations and reached general availability thirteen days later, on 9 July.
Look at what got gated. GPT-5.6 ships as Sol, the flagship for coding and cybersecurity, plus Terra and Luna. Washington slowed the flagship while Terra and Luna were unaffected. The trigger was Sol’s scores: 88.8% on Terminal-Bench 2.1 and around 96.7% on an internal cyberattack benchmark, which crossed the “High” line in the Preparedness Framework.
The hold was a gate, not a ban. Under the June executive order, the government can vet the most advanced systems for up to 30 days before release, coordinated through OSTP, ONCD and the Commerce Department. OpenAI cleared customers individually, exposing Sol only through the API and Codex. Altman described the preview as happening at Washington’s request, and said OpenAI was working to get to general availability as fast as it could. That short hold is how the pre-release access window works.
Why did OpenAI offer the US government a 5% equity stake, and what would it be worth?
OpenAI offered the US government a 5% equity stake in each leading US AI developer, worth about $42.6 billion at OpenAI’s $852 billion valuation. It framed the offer as access-plus-upside: an Alaska-style public fund channelling AI’s gains back to the public.
A thirteen-day hold does not account for a $42.6 billion concession. That mismatch is the point to watch. Sam Altman pitched the stake early with the Trump administration, under a broader arrangement where Washington would hold 5% of OpenAI, Anthropic, Google and Meta through a sovereign wealth fund. His argument: giving the public a financial interest is the best way to share AI’s upside, echoing the April “public wealth fund” proposal. There is precedent: Washington already holds about 10% of Intel and stakes in IBM. That offer is one part of the push to gate frontier models.
OpenAI publicly calls the access process “unsustainable” and says it shouldn’t become the default, yet it offered Washington ownership in the same breath. Nobody has said who holds the stake or whether Anthropic, Google and Meta would agree, which is where the securitisation of frontier AI and its vendor risk matter. While OpenAI was offering a stake, Washington was already pressing Anthropic through a different instrument: an export-control directive.
Why did Anthropic disable Fable 5 and Mythos 5, and why did it end up suspending access for everyone?
Anthropic disabled Claude Fable 5 and Mythos 5 after Amazon researchers reported a jailbreak, which prompted a 12 June Commerce export-control directive. A deemed-export clause made serving any foreign national legally equivalent to exporting the model, so a targeted response collapsed into a 19-day worldwide shutdown.
The sequence explains the “everyone” part. Fable 5 and Mythos 5 launched on 9 June on one underlying model; Fable 5 shipped with strong safeguards, Mythos 5 went only to Project Glasswing partners for defensive cybersecurity. Amazon researchers then reported a bypass of Fable 5’s safeguards, and on 12 June the Commerce Department, signed by Howard Lutnick, issued the directive on both models. Anthropic described the letter arriving at 5:21 pm ET with no stated national-security concern.
The clause that forced the worldwide shutdown was deemed export. Because a user’s nationality can’t be verified at the API layer, the only compliant move was to switch both models off for everyone. Anthropic said it had to “abruptly disable Fable 5 and Mythos 5 for all our customers”. A narrow jailbreak became a 19-day outage. Fable 5 returned globally on 1 July with an improved classifier blocking the reported bypass in over 99% of cases, while Mythos 5 stayed limited to roughly 100 vetted US organisations. The harsher treatment had a backstory: the Pentagon had already labelled Anthropic a supply-chain risk after it refused terms on autonomous weapons and mass surveillance. The shutdown is the securitisation frame in practice.
The Anthropic export-controls case vs the OpenAI GPT-5.6 case: what’s the actual difference, and how did each lab respond?
The difference is mechanism and leverage. OpenAI negotiated with equity, a 5% stake-for-access deal, while Anthropic faced an export-control directive that prohibited access outright, forcing a comply-and-litigate response. OpenAI gated Sol to about 20 trusted partners and reached general availability in 13 days; Anthropic was shut down retroactively for 19 days.
OpenAI faced a negotiable gate and answered with an instrument of its own: a stake that turned a gatekeeper into a shareholder. Anthropic faced a prohibition with no room to bargain. An export-control directive, built for chips and hardware, was stretched onto model use, and the deemed-export clause left Anthropic with compliance, litigation and a wait. One lab was compelled; the other cooperated.
The asymmetry shows in the numbers. Sol launched to about 20 vetted organisations; Mythos 5 came back to more than 100. The boundary of access was set by the mechanism. Anthropic’s path was harsher because the Pentagon had already labelled it a supply-chain risk over military-use refusals, a context OpenAI never carried. OpenAI bought itself a seat with a stake; Anthropic fought a designation from outside. That contrast is what you weigh when sizing vendor risk in the gatekeeper story.
The instrument determined the leverage, and the leverage determined the response. When Washington restricts a model you depend on, the first question is which instrument it used, because that decides how much room your provider has. For the bigger picture, see where these cases sit in the gatekeeper story.
Frequently Asked Questions
Why does OpenAI call the current government access process “unsustainable”?
OpenAI calls the process unsustainable because it asks labs to accept government veto power over releases without stable, published rules. The objection is about predictability: a system that relies on ad hoc national-security vetting, rather than clear criteria, is hard to plan a business around. The tension is that OpenAI made this complaint while simultaneously offering Washington a 5% stake in its future.
What is an export-control directive, and how can it apply to an AI model rather than hardware?
An export-control directive is a government order restricting who may receive a controlled item, and it traditionally targets physical goods such as semiconductors. Applied to AI, it restricts access to a model’s use through an API rather than the export of its weights. Commerce Secretary Howard Lutnick’s 12 June letter stretched the chip-era framework onto model use, which is why a use-based restriction could reach overseas customers.
What is a “supply chain risk designation,” and how did it shape how each lab was treated?
A supply chain risk designation is a national-security label that marks a supplier as an unacceptable dependency. The Pentagon had already applied it to Anthropic over its refusals on military use, which helps explain why the lab faced an outright prohibition rather than a negotiated gate. OpenAI carried no equivalent designation, so its path to a deal stayed open.
What is the proposed US sovereign wealth fund for AI, and who would decide how it works?
It is a proposed vehicle for holding government equity in leading US AI developers and channelling the returns back to the public, modelled on Alaska’s permanent fund. OpenAI framed its 5% offer as the first deposit. The open question is governance: the reporting does not specify who would hold the stake or who would decide on participation by Anthropic, Google and Meta.
Is the 5% equity stake a done deal, or is it still just a proposal?
It remains a proposal rather than a completed transaction. OpenAI offered Washington roughly 5% of each leading US AI developer, and the widely cited $42.6 billion figure describes what that slice would be worth at OpenAI’s reported $852 billion valuation, not cash that has changed hands. No final agreement or formal structure has been reported.
Did Anthropic choose to take Fable 5 and Mythos 5 offline, or was it compelled?
Both, in sequence. Anthropic initially disabled the models after Amazon researchers reported a jailbreak, but the worldwide suspension was compelled. The 12 June export-control directive and its deemed-export clause left no legal path to serve non-US nationals, so a targeted safeguard response became a full shutdown rather than a choice to restrict access.
Was the jailbreak that triggered the Anthropic directive actually confirmed?
Its existence was reported, but its breadth is disputed rather than confirmed. Amazon researchers reported a jailbreak of Fable 5, while Anthropic characterised the reported bypass as narrow and minor, and the administration treated it as a safety failure serious enough to justify controls. No single source adjudicates the disagreement, so the trigger remains contested even though the directive it prompted is not.
Can Washington force an AI model offline after it has already launched?
Yes. The Anthropic case is the proof: Fable 5 had already launched on 9 June when the 12 June directive forced it offline retroactively, producing the 19-day outage. The mechanism matters here, because a pre-release gate slows a launch before it happens, while an export-control prohibition can pull a live model out of service after release.
Does any of this affect companies outside the United States, such as in Australia?
Yes, and the deemed-export clause is why. It made serving any non-US national legally equivalent to exporting the model, so an Australian team using a restricted model through an API sat inside the prohibition rather than beyond it. That is what turned a targeted safeguard response into a worldwide suspension, and it is the exposure a buyer outside the US should assume.
What should a CTO watch for when building on a model the US government can gate?
Treat government access as a vendor-risk variable rather than background noise. Watch whether a provider carries a designation such as supply-chain risk, whether the instrument it faces is a negotiable gate or a flat prohibition, and how concentrated its government exposure is. The two cases show that the mechanism, not the lab’s willingness to cooperate, decides how much room a provider has.
Where can I find the primary documents on both cases?
Start with the primary documents rather than summaries. OpenAI’s GPT-5.6 Sol announcement and system card set out the release and the Preparedness Framework tier, Anthropic published statements on the Fable 5 and Mythos 5 suspension and the redeployment, and Howard Lutnick’s export-control letter is the directive itself. The April public wealth fund proposal is the other primary source behind the equity thread.
Will this happen to every future frontier model release?
Not automatically, but the possibility is now part of every flagship release. A pre-release gate is plausible whenever a model crosses a risk threshold such as the “High” cyber tier under a Preparedness Framework, while an export-control prohibition remains a more targeted instrument aimed at a specific supplier. Which mechanism Washington uses will keep determining the leverage each lab has.
