Michael C. Vincent: You came in looking like you'd already had an argument with the news.
Hope Sterling: I did have an argument with the news — I lost — because okay, I'm reading about the Anthropic ban and I'm like, wait, this is already over? It's been like two and a half weeks? I thought I missed something.
Michael C. Vincent: You didn't miss anything. That's just how fast it moved. Bureau of Industry and Security issues the order June 12 — Fable 5, Mythos 5, restricted to foreign nationals — and it's lifted inside three weeks. Shortest-lived major AI restriction on record.
Hope Sterling: And this is the first time — like, ever — that export controls got applied to an actual large language model, right? That's not me being dramatic, that's just — that's a fact?
Michael C. Vincent: That is a fact. First time in history. The Trump administration, through BIS, reached for a tool built for inspectable physical goods and aimed it at cloud software. And the whole construction lasted nineteen days.
Hope Sterling: Nineteen — stop. And the reason was because Mythos 5 was too good at finding software vulnerabilities, which — okay, that's scary, I get it, but — while all this is happening, OpenAI is literally getting a clean public launch approved for GPT-5.6 on July 9? Same administration, same window?
Michael C. Vincent: Same administration. That's what makes this genuinely strange — not just one unusual outcome, but two opposite ones running in parallel.
Hope Sterling: That's what I — okay so that's what we're actually trying to figure out today, right? Like, how did those two things happen at the same time and what does that even mean?
Michael C. Vincent: The parallel launches expose something deeper, which is that the tool they used couldn't have worked regardless of which company it targeted. Export controls are built like a checkpoint at a border crossing. You inspect the truck, you stamp the manifest, you turn away the wrong passport. Physical object, physical moment, physical enforcement. Fable 5 is not a truck. It's a webpage. There's no checkpoint.
Hope Sterling: Wait — so you can't just, like, block the URL for certain countries?
Michael C. Vincent: You can try. But Anthropic couldn't verify in real time who was a foreign national and who wasn't. So they did the only thing they could — they suspended access for everyone. Including U.S. citizens. The security measure locked out the people it was supposed to protect.
Hope Sterling: Wait, that's the part that gets me. Like, picture a cybersecurity team at an American defense contractor on June 13, they open their laptop, try to log into Fable 5, and it's just — gone. And meanwhile her peer in Toronto? Still has access. And then someone in Beijing just opens a Chinese alternative that nobody's even tracking, and —
Michael C. Vincent: That is exactly the scenario. Word for word.
Hope Sterling: So the government's attempt to protect against adversaries literally pushed the defenders toward alternatives that are — I mean, that's so much worse? CEPA actually said that, right, that the controls hurt the cyber defenders more than they hurt the adversaries?
Michael C. Vincent: CEPA's analysts made exactly that argument — that the restriction damaged U.S. credibility and handed adversaries a quieter advantage than any direct vulnerability would have. The irony being: the instrument of protection became the exposure. You see, that's not an accident of implementation. It's structural. Export controls were never designed for something that crosses every border simultaneously the moment it's deployed.
Hope Sterling: So the headline was 'government bans dangerous AI' but the actual story was 'government accidentally ran the most effective ad for switching to a Chinese alternative.' That's — I mean, that's the thing nobody led with.
Michael C. Vincent: But that framing — 'consistent call on dangerous AI' — that's the one circulating, and it doesn't survive contact with what actually happened. BIS didn't apply one standard to two companies. Anthropic got a blanket public ban. Fable 5, Mythos 5, gone for everyone. OpenAI, same window, got a supervised preview of GPT-5.6 through Codex with a vetted partner set.
Hope Sterling: Wait — through Codex specifically? Like, not even ChatGPT yet, just the coding platform?
Michael C. Vincent: Codex. Trusted partners, federal oversight. Which sounds rigorous until you ask — what criteria put one company in the 'hammer' column and the other in the 'supervised preview' column? Because no transparent criteria were published. None.
Hope Sterling: Okay but — wait, that's the part that's messing with me, because like, if the security concern was Mythos 5's ability to find exploitable vulnerabilities, GPT-5.6 Sol is also a frontier model designed for — what did they call it — complex reasoning and agentic workloads? So why does one get a ban and the other gets a launch date of July 9 fully approved? That's not a consistent standard, that's a — I mean, that's a negotiation outcome dressed up as a security ruling.
Michael C. Vincent: And now look at what OpenAI walked into that window with. Sol at five dollars input, thirty dollars output per million tokens. Terra at two-fifty and fifteen. Luna at one dollar and six. That is not a model — that is a fully staged product line. Three tiers, three price points, ready to deploy the moment the preview ended.
Hope Sterling: No way — that level of pricing architecture doesn't get built during a three-week ban. That's months of prep.
Michael C. Vincent: Which is exactly the discomfort the record leaves you with. Not an accusation — but the question of whether the sequencing was coincidence or something more coordinated is, well, the question neither the Trump administration nor OpenAI has answered.
Hope Sterling: And — okay, I'm not saying anyone planned this, but like, if you're a competitor watching Anthropic get hammered while OpenAI gets a glide path to the biggest launch of the year? That's not a policy. That's a market outcome. And honestly the geopolitical piece of this — what China actually gains every time U.S. companies are disrupted in the name of security — that's the part that comes later and makes this so much harder to dismiss.
Michael C. Vincent: You see, that's the thread worth pulling. Because what the record shows isn't malice — it's something arguably more unsettling. A policy that was ad hoc enough to produce two opposite outcomes for two companies in the same fortnight, with no published explanation for why.
Hope Sterling: And that's the thing — every day the controls were on, China wasn't sitting still. No self-imposed pause, no BIS directive, nothing. Just — open field.
Michael C. Vincent: That is the exact asymmetry. Every period of U.S. market disruption is a period of uncontested Chinese expansion. And CEPA named this specifically — the credibility damage isn't just reputational. It is a concrete signal to allied governments that American frontier AI access is a political variable. Not infrastructure. A variable.
Hope Sterling: Wait — allied governments. Like, European partners had to watch this and think — oh, our entire stack could just... disappear?
Michael C. Vincent: Without warning. That's the phrase worth sitting with. A ministry of defense in Warsaw, say, mid-procurement on an AI-assisted threat analysis tool built on an American provider — June 12 happens, and they now know: this can be revoked. No notice, no transition window.
Hope Sterling: And then July 9 arrives and — GPT-5.6 Sol, Terra, Luna, the whole product line drops into ChatGPT and the partner ecosystem on the same day Anthropic's restrictions lift? That timing is — I mean, that's not accidental, that's — wait, actually, is OpenAI's integration being already baked in the real story here?
Michael C. Vincent: That is the story. OpenAI lands that moment with Sol, Terra, and Luna already threaded into ChatGPT, already in the partner ecosystem. Anthropic restarts from a cold stop. The simultaneous availability is a reset in name only — OpenAI's architecture never left.
Hope Sterling: So the managed Codex preview wasn't just a workaround — it was a runway.
Michael C. Vincent: A learned behavior, I'd say. The Codex preview mirrored exactly the trusted-partner framework BIS itself uses for sensitive export decisions. Whether that was deliberate or intuitive — well, the effect is the same. Government coordination as a competitive moat. If that pattern normalizes, the approval pipeline becomes the barrier to entry. Incumbents are entrenched. Challengers — domestic or foreign — wait.
Hope Sterling: And nobody voted for a regulatory moat. That's just — that's what we inherited from nineteen days.
Michael C. Vincent: And that's the question that doesn't resolve. Six months from now — does the Commerce Department publish actual written criteria for which frontier models get a blanket ban and which get a supervised preview? Or does every case stay a back-room negotiation with no published standard?
Hope Sterling: Because right now there's literally nothing — no transparent criteria for why Anthropic's models got the hammer and OpenAI got Codex and a runway. And China has zero equivalent self-imposed restrictions on its frontier developers. Not one. So if the answer is 'no framework,' we just — I mean, what did we actually do here?
Michael C. Vincent: You managed scarcity. Dressed it as security. And left the field open.
Hope Sterling: Yeah. That one's going to sit with me for a while.
