Jonathan Ingles: Two frontier labs. One day. Both gated by the U.S. government. Axios described it as the first known instance of the government preemptively restricting a domestic AI model launch. That's the sentence. June 26th, 2026.
Maya Chen: Hold on.
Jonathan Ingles: OpenAI previews GPT-5.6 — Sol, Terra, Luna — to approximately twenty government-approved partners. Not the public. Not developers. Twenty partners the government cleared. ONCD and OSTP delivered that request straight to Sam Altman.
Maya Chen: And it's not — I mean, it's not even like a quiet administrative thing. Altman told staff internally. The Information had the memos. This is the CEO explaining to his own people why their product launch got... handed to a government approval queue.
Jonathan Ingles: And Anthropic. Same day. Mythos 5 cleared for a hundred organizations. Fable 5 — same company, hit by the same export-control directive around June 12th — still completely offline. No explanation.
Maya Chen: Yeah, that's the part that sort of — the way I'd put it to anyone who hasn't been following this closely: it's like your phone carrier needing White House sign-off to activate your new service. Except two carriers, same afternoon, and one of them only got half their phones turned back on.
Maya Chen: But wait — that's actually what I want to land on. Because government having opinions about AI, that's not new. The thing that's new here is that ONCD and OSTP didn't comment after the fact. They acted before. Preemptive veto. No public threat assessment. Nothing released. We're just supposed to accept that the cybersecurity rationale exists somewhere classified.
Jonathan Ingles: And OpenAI basically confirmed the duress. Publicly.
Maya Chen: Right — they said they don't believe this process should become the long-term default. That's not a company at peace with this. You don't say 'this shouldn't be permanent' unless you're trying to signal something to someone. Your users, your investors, someone.
Jonathan Ingles: And this is the Trump administration. The one that ran on deregulation. Howard Lutnick's in the room on the approval side. That reversal has never been explained.
Maya Chen: No, and — okay, actually here's what I find harder to ignore than even that. Fable 5 is still completely dark. Mythos 5 gets a hundred organizations. Same company, same export-control directive, two different outcomes. Business Insider flagged Anthropic as potentially the largest loser here, while simultaneously they got the carve-out. That contradiction — that's not policy. That's leverage.
Jonathan Ingles: The gap between the two models is the whole story.
Jonathan Ingles: The circulating take is that this is a cybersecurity measure. Kill it. If it were, DeepSeek faces zero equivalent restrictions. Zero. A Chinese competitor ships freely while ONCD and OSTP gate GPT-5.6. That's not a security posture. That's a selection.
Maya Chen: Hold on — so the gate doesn't reduce the capability risk in the world at all.
Jonathan Ingles: It moves access to foreign models. That's the actual consequence. A grad student at UC Davis, lab not on the approved partner list, supposed to start a medical imaging project on GPT-5.6 — she can't. DeepSeek's latest ships without restriction, benchmarks faster anyway. By choosing, she's handed the access advantage to a lab the government hasn't touched.
Maya Chen: Okay but — wait, no, I want to press on this. Is the answer remove the gate, or extend it globally? Because those are completely different problems and I'm not sure which one we're actually arguing.
Jonathan Ingles: Neither matters until there's a public threat assessment. There isn't one. You can't evaluate proportionality on a classified rationale. That's not safety policy.
Maya Chen: And who decides when it's proportionate? Because right now the gate structurally advantages OpenAI and Anthropic — they have compliance infrastructure, Washington relationships. The researcher without a government relations officer is just... out.
Jonathan Ingles: That's leverage, not policy. No transparency, no proportionality check — the cybersecurity framing is the thing nobody's supposed to question. That's the tell.
Maya Chen: And that's the part I keep sitting with, actually — not whether the gate is justified right now, but... once OpenAI has gone on record saying this shouldn't be the long-term default while cooperating with it anyway, that cooperation is the norm. That's the thing. The stated intent doesn't matter if the structure is already in place. Future administrations inherit the approved-partner list. They don't inherit the asterisk.
Jonathan Ingles: And foreign governments watch it work. That's the export. Not GPT-5.6. The model — the governance model. ONCD and OSTP demonstrated that you can preemptively gate a frontier lab with a classified rationale and the lab complies. That playbook just got published.
Maya Chen: Which is — yeah. So the question that I don't think we've actually answered is: what's the political incentive for any administration to ever delete the approved-partner list once it exists?
