Mark Delaney: Hey, good week? — actually, doesn't matter, because I found the thing that's been living in my head rent-free and I need to just say it out loud.
Michael C. Vincent: Go on.
Mark Delaney: The Trump administration held GPT-5.6 hostage for thirteen days, and then just... let it go. No explanation. And I can't figure out if that means the threat was fake or the deal was real.
Michael C. Vincent: Now, that is the question. June 26, 2026 — OpenAI releases GPT-5.6, three tiers: Sol, Terra, Luna. Flagship down to cost-efficient. And the whole family is gated, restricted to vetted partners at the explicit request of the U.S. government. National-security concerns, they said.
Mark Delaney: Wait — explicit request. Like, OpenAI agreed to this?
Michael C. Vincent: Sam Altman told his staff the government would be 'approving access customer by customer.' That's the phrase. And Howard Lutnik — the Commerce Secretary — was in direct conversation with Altman about the model. That's not OpenAI pushing back. That's OpenAI at the table.
Mark Delaney: And Anthropic went through basically the same thing right before this, uh — Fable 5 and Mythos 5, export controls imposed, then lifted in under three weeks. Same pattern. So the real question is whether any of this was ever actually about China, or whether it was about who controls the dial.
Michael C. Vincent: That is precisely the arc today. Not whether GPT-5.6 posed a risk. Whether any democracy can govern transformative power when it is terrified and improvising simultaneously.
Mark Delaney: But wait — before we get to who controls the dial, I want to nail down what the threat actually was, because I think we skipped that. Like, the DMV holds your car keys for two weeks, then hands them back, and never says what they checked. That's the whole thing. What was the check?
Michael C. Vincent: The stated threat was China. Every time. Adversarial access to frontier capability.
Mark Delaney: Right — but the part that doesn't fit is Mythos 5. Because with Anthropic's model, they got specific. The government flagged it for a reported ability to find and exploit software vulnerabilities. That's not vague. That's a named capability. And then Dario Amodei comes out around June 10th and basically says — uh, no, that's overstated. Anthropic disputed their own model's threat characterization.
Michael C. Vincent: I did not see that coming — the company telling the government its own risk assessment was wrong.
Mark Delaney: And then CEPA — they published on July 7th, 2026 — they argued the Mythos export controls were misdirected, risk overstated, and that the whole episode actually damaged U.S. credibility as a reliable partner. So you've got the lab disputing it, you've got outside analysts disputing it, and the government never explained what changed when they lifted it. Neither reversal — not Anthropic's, not GPT-5.6 — came with any public accounting.
Michael C. Vincent: You see, that silence is the tell.
Mark Delaney: I mean — maybe. Or, wait, actually I want to be careful here, because we don't know what was in any classified review. There could be something real we just can't see. I'm not saying it was pure theater. I'm saying the public case collapsed and they never rebuilt it out loud.
Michael C. Vincent: The evidence they showed us — and 'showed' is generous — didn't survive the scrutiny of the people who built the model. That's the core of it.
Mark Delaney: And that's actually where the no-law thing bites hardest. Because if there's no statute, there's no criteria for what 'the public case' even has to look like. Trump signed an executive order — the AI cybersecurity clearinghouse at Treasury, the 30-day frontier model review — but that's an executive order. That's not a law. That's a president telling agencies what to do until the next president tells them something different.
Michael C. Vincent: And this is the part worth pausing on. The same administration had *previously* rejected government AI reviews. That was the position. Then they mandated a 30-day review process. No explanation for the reversal. The executive order is the about-face.
Mark Delaney: Wait — they flipped their own stated position?
Michael C. Vincent: Without a published explanation, yes. And there are no criteria — none that NIST published, none that Commerce posted. So the customer-by-customer approval for GPT-5.6 wasn't administering a framework. It was the government picking winners.
Mark Delaney: Okay, so — uh, picture this. You're a healthcare startup. You've been pre-approved for GPT-5.6 access, you've got investor meetings booked, engineers on payroll. June 26th hits and instead of an API key you get a call: government review, customer by customer, no timeline. No one can tell your VP of engineering whether this resolves in three days or three months. And two weeks later the whole thing lifts — nothing. No letter, no criterion met, nothing. How do you even explain that to your board?
Michael C. Vincent: You can't. And that's the regulatory capture problem sitting right underneath it. The U.S. Department of Commerce is deciding, customer by customer, which commercial actors get early access to OpenAI's most powerful models. With no published standard. That's not security architecture. That's a gatekeeper.
Mark Delaney: And whoever's well-connected in Washington gets the call returned faster.
Michael C. Vincent: That is the kernel. Right there. The hot take has that much exactly right — there is no law, no criteria, no framework. Just Howard Lutnik on the phone with Sam Altman. Executive muscle dressed as policy.
Mark Delaney: And the China rationale — I want to just flag, because we're going to come back to it — there's a real argument that the way they're using China actually undermines the policy they're claiming to run.
Michael C. Vincent: And that's where it falls apart on its own terms. Because China is the named threat — every briefing, every justification. But if China already has comparable frontier models, what does a thirteen-day delay on U.S. public access actually buy?
Mark Delaney: Nothing. It buys nothing from a strategic standpoint. But it does buy something for OpenAI and Anthropic — government-adjacent positioning, a kind of... uh, legitimacy stamp. Like, Commerce cleared us. That's not nothing commercially.
Michael C. Vincent: Incumbent entrenchment.
Mark Delaney: Exactly — and that's the part that actually worries me more than the China framing. Because the policy claims to defend the broad U.S. AI ecosystem, right? That's the stated goal. But who got the early access? Who had direct lines to Howard Lutnik? OpenAI. Anthropic. Not the startup in Pittsburgh, not a university lab. So the Department of Commerce ran a process that — I mean, whether they meant to or not — it concentrated access at the top.
Michael C. Vincent: And allies watching this — what do they see?
Mark Delaney: An unreliable partner. That's the CEPA read from July 7th, and I think they're right. If you're a European firm or a Japanese lab trying to build on U.S. frontier models, you just watched the U.S. government freeze access without criteria, lift it without explanation, and do it twice — Fable 5, Mythos 5, then GPT-5.6. You're not thinking 'strong security framework.' You're thinking — do I build my stack on this?
Michael C. Vincent: The threat rationale, when you apply pressure to it, actually accelerates the thing it claims to prevent.
Mark Delaney: That's the calibrated version, I think. Not pure hostage theater — I'll grant that maybe the classified review caught something real. We genuinely can't know. But the public logic is self-defeating: name China as the threat, deliver no strategic delay that matters to China, damage allied trust, and hand incumbents a moat. The Department of Commerce cleared GPT-5.6 on July 9th with no published criteria. That's not a security outcome. That's an improvisation that will get cited as precedent.
Michael C. Vincent: And the next lab that goes through this process will have no more clarity on the rules than OpenAI did.
Mark Delaney: You know what's funny — we started this whole thing with 'the Trump administration held GPT-5.6 hostage for thirteen days and then just let it go.' And I think where I land is: it wasn't a hostage situation. It was something almost weirder. It was a proof of concept with no written-down criteria, and now it exists as precedent. July 9th, 2026 — cleared. No public account of what the review found, no explanation of what changed. That's not a hostage release. That's a tool that got tested and put back in the drawer.
Michael C. Vincent: And the next administration — any administration — opens that drawer and finds a two-week proof of concept that it works. No statute. No NIST framework. No published standard at Commerce. Just the memory that it happened and nobody stopped it.
Mark Delaney: That's — yeah. The most dangerous thing wasn't restricting GPT-5.6. It was proving you can, and no one has to explain why.
Michael C. Vincent: Well. That'll sit with me for a while.
Mark Delaney: Same. Good one to chew on.
